Category Archives: Observability

AIOps – Artificial Intelligence for IT Operations

AIOps stands for Artificial Intelligence for IT Operations. It is a methodology for using machine learning, statistical analysis, automation, and now LLM-based reasoning to improve how infrastructure and application operations teams detect, investigate, explain, and resolve problems.

At its core, AIOps is trying to solve a very practical problem:

Modern systems produce more operational data than humans can manually inspect, correlate, and act on quickly enough.

That includes metrics, logs, traces, events, alerts, tickets, deployments, topology changes, CI/CD activity, cloud audit logs, Kubernetes events, OpenStack state, Slurm queues, GPU telemetry, network flows, and user-impact signals.


1. The Problem AIOps Is Trying to Solve

Modern IT operations has become too complex for purely manual troubleshooting.

A typical platform may include:

Users

Load balancers

Ingress / API gateways

Kubernetes services

Microservices

Databases / queues / object storage

Cloud / OpenStack / VMware / bare metal

Networks / firewalls / DNS / storage / GPUs

Every layer emits telemetry. The problem is not lack of data. The problem is too much disconnected data.


Problem 1: Alert Fatigue

Operations teams often receive hundreds or thousands of alerts.

Many are:

  • duplicates
  • symptoms rather than root causes
  • low priority
  • transient
  • missing context
  • caused by the same underlying event

Example:

Disk latency high
API latency high
Pod restart count high
Database connection errors
Frontend 500s
SLO burn rate alert
User complaints

A human has to determine whether these are six separate incidents or one cascading failure.

AIOps tries to group these signals into one meaningful incident.


Problem 2: Data Silos

Metrics are in one place.

Logs are in another.

Traces are somewhere else.

Tickets are in Jira or ServiceNow.

Deployments are in GitLab or GitHub.

Infrastructure state is in OpenStack, Kubernetes, Slurm, Ceph, AWS, Azure, or VMware.

The engineer has to jump between tools:

Grafana → Loki → Tempo → Prometheus → Kubernetes → OpenStack → SSH → Jira → GitLab

That is slow, error-prone, and dependent on tribal knowledge.

AIOps tries to connect these sources and reason across them.


Problem 3: Manual Root Cause Analysis

Traditional troubleshooting is often manual correlation.

An engineer asks:

What changed?
What broke?
Who deployed?
Which node is affected?
Is this network, storage, compute, DNS, auth, GPU, database, or app?
Has this happened before?
What fixed it last time?

That investigation may take 30 minutes, 2 hours, or several days.

AIOps attempts to reduce that investigation time by automatically correlating evidence.


Problem 4: Too Much Complexity

Modern platforms are dynamic.

Examples:

  • containers are rescheduled
  • pods are ephemeral
  • cloud instances appear and disappear
  • autoscaling changes capacity
  • CI/CD continuously deploys changes
  • service dependencies shift
  • storage volumes move
  • GPU nodes are drained, allocated, or isolated
  • network paths change
  • certificates expire
  • DNS records update

Humans are not good at mentally tracking all of that in real time.

AIOps tries to build a continuously updated operational view of the environment.


Problem 5: Reactive Operations

Traditional operations is often reactive:

Something breaks → alert fires → engineer investigates → fix applied

AIOps aims to make operations more proactive:

Early warning → anomaly detected → likely cause identified → risk predicted → action recommended

For example:

  • predict disk exhaustion
  • detect memory leak patterns
  • identify increasing error budgets burn
  • spot noisy neighbours
  • detect degraded GPU nodes
  • find abnormal API latency before customers complain
  • flag risky deployments

2. What AIOps Is Trying to Solve

AIOps is trying to make IT operations:

  • faster
  • more accurate
  • less noisy
  • more automated
  • more predictive
  • less dependent on individual experts
  • better aligned with business and service impact

The goal is not simply “AI for dashboards.” The goal is to improve operational outcomes.


3. The Main Capabilities of AIOps

A mature AIOps approach usually includes several capabilities.


1. Data Collection

AIOps needs telemetry from across the environment.

Typical sources include:

Metrics       → Prometheus, Mimir, VictoriaMetrics, CloudWatch
Logs → Loki, Elasticsearch, OpenSearch, Splunk
Traces → Tempo, Jaeger, OpenTelemetry
Events → Kubernetes events, OpenStack events, systemd, audit logs
Tickets → Jira, ServiceNow, GitLab issues
Deployments → GitLab CI/CD, GitHub Actions, ArgoCD
Infrastructure→ OpenStack, Kubernetes, Ceph, Slurm, VMware, AWS, Azure
Network → flow logs, DNS logs, firewall logs, load balancer logs

Without good data, AIOps is weak. The first requirement is solid observability.


2. Noise Reduction

AIOps should reduce alert noise by grouping related alerts.

For example, instead of showing:

NodeDown
PodCrashLooping
HTTP5xxHigh
LatencyHigh
DatabaseConnectionFailure
SLOBurnRateHigh

it should produce something closer to:

Incident: Database node failure causing API errors

Affected services:
- checkout-api
- payment-api
- frontend

Likely root cause:
- PostgreSQL primary unavailable

Evidence:
- node db-03 stopped responding at 10:42
- API connection errors started at 10:43
- customer-facing 500s increased at 10:44

This is one of the biggest practical wins of AIOps.


3. Anomaly Detection

AIOps can learn normal behaviour and detect deviations.

Examples:

CPU usage normally peaks at 70%, now 95%
API latency usually 120 ms, now 900 ms
GPU memory errors normally zero, now increasing
Login failures normally 20/hour, now 5,000/hour
Network packet drops normally rare, now concentrated on one host

This is useful when static thresholds are poor.

A static alert might say:

CPU > 90%

But anomaly detection can say:

This service normally uses 15% CPU at this time of day.
It is now using 65%, which is abnormal for this workload.

That is more context-aware.


4. Event Correlation

AIOps correlates events across systems.

Example:

10:01 - GitLab deployment completed
10:04 - Kubernetes pods restarted
10:05 - latency increased
10:06 - error rate increased
10:07 - SLO burn alert fired

The likely cause is not “latency high.” The likely cause is the deployment.

AIOps should connect those facts.


5. Root Cause Analysis

AIOps tries to identify the underlying cause, not just the symptoms.

Example:

Symptom:
Users cannot access the application.

Possible causes:
- DNS failure
- certificate expiry
- ingress failure
- pod crash
- database outage
- network ACL issue
- storage outage
- failed deployment

AIOps task:
Rank the most likely causes using evidence.

A good AIOps system does not just say:

Application is down.

It says:

The application is down because the ingress controller cannot reach the backend pods.
The backend pods are healthy, but the service selector was changed in the latest deployment.

That is operationally useful.


6. Recommendation

AIOps should recommend next actions.

Example:

Recommended action:
Rollback deployment checkout-api:v2.4.1 to v2.4.0.

Reason:
Errors started within 3 minutes of the deployment.
No infrastructure errors were detected.
Previous version had normal latency and error rate.

The recommendation should include evidence, not just a guess.


7. Automation and Remediation

At higher maturity, AIOps can automate approved actions.

Examples:

Restart a failed service
Scale a deployment
Rollback a release
Drain a bad Kubernetes node
Evacuate an OpenStack compute node
Restart a failed exporter
Open a Jira ticket
Page the correct team
Run a known Ansible playbook

But automation should be controlled carefully.

The safest path is usually:

Detect → Explain → Recommend → Human approval → Execute → Verify

Only mature, low-risk, well-tested actions should be fully automatic.


4. AIOps Compared With Traditional Observability

Traditional observability answers:

What is happening?

AIOps tries to answer:

What is happening?
Why is it happening?
What changed?
What is the blast radius?
What should we do next?
Can we fix it automatically?

Observability provides the evidence.

AIOps provides correlation, reasoning, prioritisation, and action.

They are not competitors. AIOps depends on observability.


5. AIOps in Your MCP Example

In the previous MCP example, the user asks:

Why did gpu-test-01 fail to start?

A traditional engineer might manually check:

openstack server show gpu-test-01
openstack console log show gpu-test-01
openstack port list
openstack hypervisor list
docker logs nova_scheduler
docker logs nova_compute
journalctl on compute nodes
sinfo
nvidia-smi
kubectl get pods
Grafana dashboards
Loki logs
Prometheus GPU metrics

An MCP-enabled AIOps agent could do much of this automatically.

It could query:

OpenStack MCP   → VM state, scheduler errors, Neutron ports
Nova MCP → compute scheduling failure
Neutron MCP → network binding or DHCP issue
Slurm MCP → GPU node allocation or drain state
Prometheus MCP → CPU, RAM, disk, GPU health
Loki MCP → Nova, libvirt, Neutron logs
Kubernetes MCP → GPU Operator / NVIDIA plugin status
Ceph MCP → storage availability
Ansible MCP → known remediation playbooks

Then return something useful:

gpu-test-01 failed because Nova could not schedule the requested PCI device.
The requested alias nvidia-gpu-audio is not defined in nova.conf.
The VM requested a GPU-related PCI alias that the scheduler cannot match.

Evidence:
- Nova API returned PCI alias nvidia-gpu-audio is not defined
- No matching pci_alias exists on the compute configuration
- Hypervisor is otherwise healthy
- Neutron port exists
- Image and flavor are valid

Recommended fix:
Add or remove the correct PCI alias definition, reconfigure Nova, restart nova-scheduler and nova-compute, then retry the server create command.

That is AIOps because the system has moved beyond raw monitoring and into assisted diagnosis.


6. The AIOps Methodology

AIOps is not just a product. It is a way of operating.

A practical methodology looks like this:

1. Instrument everything
2. Centralise telemetry
3. Normalise and enrich the data
4. Correlate events across systems
5. Detect anomalies
6. Identify service impact
7. Recommend actions
8. Automate safe remediation
9. Verify outcomes
10. Learn from incidents

The goal is continuous operational learning.

Every incident should improve the system.


7. The Maturity Model

AIOps adoption usually happens in stages.


Level 1 — Better Visibility

You collect metrics, logs, traces, and events.

Typical tools:

Prometheus
Grafana
Loki
Tempo
OpenTelemetry
Elasticsearch
Alertmanager

At this level, humans still do most of the reasoning.


Level 2 — Alert Correlation

You start grouping alerts into incidents.

Example:

20 alerts → 1 incident

This reduces noise and improves response time.


Level 3 — Assisted Investigation

The system helps engineers investigate.

It can answer:

What changed?
What services are affected?
Are there similar previous incidents?
Which logs matter?
Which deployment caused this?

This is where LLMs and MCP become very useful.


Level 4 — Recommendation

The system recommends fixes.

Example:

Rollback service X
Restart exporter Y
Scale deployment Z
Drain node A
Check Ceph OSD B
Renew certificate C

Humans still approve the action.


Level 5 — Automated Remediation

The system performs low-risk actions automatically.

Example:

Restart crashed exporter
Re-run failed health check
Scale stateless service
Create incident ticket
Attach logs and traces
Notify owning team

High-risk actions still require approval.


8. Who Should Adopt AIOps?

AIOps is most valuable for teams running complex, distributed, high-volume, or business-critical systems.


1. SRE Teams

SRE teams are one of the best fits.

They already care about:

SLIs
SLOs
error budgets
incident response
toil reduction
automation
reliability engineering

AIOps helps SREs reduce repetitive investigation and focus on higher-value engineering.


2. Platform Engineering Teams

Platform teams should adopt AIOps when they operate shared platforms such as:

Kubernetes
OpenStack
VMware
Ceph
GitLab
ArgoCD
CI/CD platforms
internal developer platforms

AIOps helps them understand platform-wide impact and detect shared infrastructure failures.


3. Observability Teams

Observability engineers are central to AIOps.

They provide the telemetry foundation:

metrics
logs
traces
events
dashboards
alerts
instrumentation standards
OpenTelemetry pipelines

Without observability engineering, AIOps becomes guesswork.


4. NOC Teams

Network Operations Centres can use AIOps to reduce noise and improve triage.

Common use cases:

deduplicating network alerts
identifying link degradation
correlating firewall, DNS, BGP, and load balancer events
detecting regional outages
routing incidents to the right team

5. Cloud Infrastructure Teams

Teams running cloud platforms benefit heavily.

Examples:

OpenStack private cloud
AWS landing zones
Azure platforms
GCP platforms
hybrid cloud
multi-cloud

AIOps can correlate compute, network, storage, identity, quota, and deployment issues.


6. HPC and GPU Platform Teams

This is especially relevant for AI infrastructure.

A GPU/HPC platform has many failure domains:

GPU health
PCI passthrough
NVIDIA drivers
CUDA versions
Slurm queues
Kubernetes GPU Operator
RDMA / RoCE
InfiniBand
Ceph / Lustre / WEKA / DDN
container runtimes
job scheduling
tenant quotas
power and thermal limits

AIOps can help detect degraded GPU nodes, scheduling bottlenecks, failed jobs, and noisy tenants.


7. DevOps Teams

DevOps teams can use AIOps to connect deployment activity to runtime impact.

Example:

Deployment happened → error rate increased → SLO burn increased → rollback recommended

This is especially valuable in CI/CD-heavy environments.


8. Enterprises With 24/7 Services

Any organisation with critical always-on services should consider AIOps.

Examples:

financial services
telecoms
cloud providers
SaaS companies
healthcare platforms
universities
government services
e-commerce
AI infrastructure providers

The more expensive downtime is, the more valuable AIOps becomes.


9. Who Does Not Need Heavy AIOps Yet?

AIOps may be overkill for very small or simple environments.

For example:

one small website
one database
low traffic
few alerts
manual operations are still manageable
no 24/7 support requirement

These teams should first focus on:

basic monitoring
good backups
clear alerts
simple runbooks
patching
logging
uptime checks

AIOps should not be used to compensate for weak fundamentals.


10. What AIOps Requires Before It Works Well

AIOps needs a strong foundation.


Good Telemetry

You need reliable metrics, logs, traces, and events.

Bad data produces bad recommendations.


Good Service Ownership

The system must know:

who owns the service
who is on call
what the service depends on
what its SLO is
where the runbook is

Without ownership metadata, routing and remediation are weak.


Good Topology

AIOps needs to understand relationships.

Example:

frontend depends on checkout-api
checkout-api depends on postgres
postgres runs on node db-03
db-03 uses ceph-volume-17
ceph-volume-17 depends on osd-4
osd-4 runs on storage-node-2

Topology allows the system to understand blast radius.


Good Change Data

Most incidents are caused by change.

AIOps should ingest:

deployments
config changes
Terraform changes
Ansible runs
package upgrades
Kubernetes rollouts
OpenStack reconfigures
firewall changes
DNS changes
certificate renewals

Without change data, root cause analysis is incomplete.


Good Runbooks

AIOps automation depends on safe, tested actions.

Examples:

restart service
rollback deployment
clear failed job
rotate certificate
drain node
restart exporter
scale deployment
fail over service

If the runbooks are poor, automation becomes dangerous.


11. Risks and Anti-Patterns

AIOps can fail if implemented badly.


Risk 1: Treating AIOps as Magic

AIOps is not magic.

It cannot fix poor monitoring, poor architecture, missing logs, or unclear ownership.


Risk 2: Automating Too Soon

Do not let AI perform destructive actions before trust is established.

Dangerous actions include:

delete data
restart databases
modify firewall rules
change identity policies
drain production clusters
detach storage
scale expensive GPU workloads

Start with read-only analysis, then human-approved remediation.


Risk 3: Poor Explainability

AIOps must explain why it thinks something is wrong.

Bad:

Root cause: database.

Good:

Root cause is likely PostgreSQL primary saturation.
Evidence:
- connections reached max at 10:42
- API errors began at 10:43
- no deployment occurred
- CPU and disk IO increased on db-01
- similar incident occurred last month

Operations teams need evidence, not vague AI output.


Risk 4: No Human Governance

AIOps should respect operational controls:

approval workflows
audit logs
change windows
RBAC
break-glass access
compliance boundaries
incident commander authority

The AI should support operations, not bypass them.


12. What Success Looks Like

A successful AIOps implementation should improve measurable outcomes.

You should track:

MTTA  - mean time to acknowledge
MTTR - mean time to resolve
MTTD - mean time to detect
alert volume
false positive rate
incident recurrence
toil hours
escalation rate
SLO compliance
change failure rate
automation success rate

The goal is not “we added AI.”

The goal is:

fewer noisy alerts
faster diagnosis
better root cause analysis
lower toil
higher service reliability
safer automation
more consistent operations

13. Practical Adoption Path

A good adoption path would be:

Phase 1: Centralise telemetry
Phase 2: Improve alert quality
Phase 3: Add service ownership and topology
Phase 4: Correlate events and changes
Phase 5: Introduce AI-assisted investigation
Phase 6: Add recommendation workflows
Phase 7: Automate low-risk remediation
Phase 8: Continuously review incidents and improve models/runbooks

For your kind of environment, the most natural starting point would be:

Prometheus/Mimir + Loki + Tempo + OpenTelemetry

Service and infrastructure inventory

Alert correlation

LLM/MCP assistant for investigation

Human-approved Ansible remediation

Closed-loop AIOps

Bottom Line

AIOps is a methodology for making operations smarter by combining:

observability
event correlation
machine learning
LLM reasoning
automation
service ownership
incident management
runbooks
governance

It is trying to solve the operational overload caused by modern distributed systems.

The teams that should adopt it first are:

SRE teams
platform engineering teams
observability teams
cloud infrastructure teams
NOC teams
DevOps teams
HPC/GPU platform teams
enterprises running critical 24/7 services

The best way to think about it is:

Observability tells you what happened.
AIOps helps explain why it happened, what it affects, and what to do next.

Observability Advances for Effective AIOps

Observability is arguably the most important technical component of AIOps.

AIOps is only as good as the operational data it can reason over. The AI layer does not magically understand your systems; it needs evidence. That evidence comes mainly from observability.

You can think of AIOps like this:

AIOps = AI reasoning + Observability data + Automation + ITSM/process + Governance

Or more practically:

Observability provides the evidence.
AI performs correlation and reasoning.
Automation executes safe actions.
ITSM/process manages incidents and ownership.
Governance keeps it controlled and auditable.

Why observability is central

Observability gives the AIOps system the raw material it needs:

Metrics  → What is slow, saturated, failing, or abnormal?
Logs → What actually happened inside the system?
Traces → Where did the request slow down or fail?
Events → What changed in the platform?
Alerts → What conditions crossed operational thresholds?
Topology → What depends on what?

Without this, AI has no reliable basis for diagnosis.

For example, if an AI agent is asked:

Why did gpu-test-01 fail to start?

It needs observability and operational signals from:

OpenStack state
Nova scheduler logs
Neutron events
Libvirt errors
Prometheus metrics
Loki logs
Slurm node state
GPU telemetry
Kubernetes events
Ceph health
Recent Ansible or config changes

The AI then correlates those signals into a root-cause explanation.

The hierarchy of AIOps components

I would rank the major components like this:

1. AI / ML reasoning layer

This includes:

anomaly detection
event correlation
root cause analysis
prediction
recommendation
LLM-based investigation

This is the “intelligence” part.

2. Observability and telemetry

This is the evidence layer:

metrics
logs
traces
events
alerts
service health
infrastructure state
change data

This is probably the most important foundation.

3. Topology and context

AIOps needs to know relationships:

service A depends on service B
pod runs on node X
node X uses storage volume Y
volume Y depends on Ceph OSD Z
tenant workload uses GPU node N

Without topology, the AI may see symptoms but struggle to understand blast radius.

4. Automation and remediation

This turns insight into action:

restart service
rollback deployment
scale workload
drain node
run Ansible playbook
open incident ticket
notify owning team

This is where AIOps becomes operationally powerful.

5. ITSM / incident workflow

This includes:

Jira
ServiceNow
PagerDuty
Opsgenie
Slack
Teams
runbooks
postmortems
ownership metadata

This connects AIOps to how humans actually operate.

6. Governance and security

Especially important if the system can take action:

RBAC
approval gates
audit logs
change windows
human approval
secrets control
tool permissions

So the answer is: yes, but with nuance

Yes — observability is the major foundation after AI.

But in a mature AIOps system, observability alone is not enough. You also need:

good topology
good ownership data
good change tracking
good runbooks
safe automation
clear governance

The best way to phrase it is:

AI is the reasoning engine of AIOps.
Observability is the evidence layer.
Automation is the action layer.
ITSM and governance are the control layer.

For your homelab and OpenStack/GPU/Slurm/Kubernetes direction, the key AIOps foundation would be:

Prometheus/Mimir  → metrics
Loki → logs
Tempo → traces
OpenTelemetry → instrumentation
Alertmanager → alert routing
Grafana → visualisation
MCP → AI tool access
Ansible → controlled remediation
GitLab → change/deployment context

Describe the advances in Observability to make AIOps effectiveThought for 34s

Yes. The big advances in observability that make AIOps genuinely effective are not just “more dashboards.” They are about turning raw telemetry into structured, correlated, machine-readable operational evidence that AI can reason over.

The shift is:

Old monitoring:
Thresholds, dashboards, isolated alerts, manual diagnosis.

Modern observability for AIOps:
Correlated metrics, logs, traces, profiles, events, topology, ownership,
change history, and automation context.

AIOps needs observability to answer five operational questions:

What is happening?
Where is it happening?
Why is it happening?
What changed?
What should we do next?

1. Standardised Telemetry: OpenTelemetry

One of the biggest advances is OpenTelemetry.

Historically, every observability vendor or tool had its own agent, SDK, format, and metadata conventions. That made AIOps hard because the AI had to reason across inconsistent data.

OpenTelemetry helps by giving teams a vendor-neutral way to instrument, generate, collect, and export telemetry such as traces, metrics, and logs. Its Collector provides a common way to receive, process, and export telemetry, reducing the need to run many different agents.

For AIOps, this matters because AI performs better when telemetry has consistent structure.

Example:

Bad:
"error happened on server"

Better:
service.name=checkout-api
deployment.environment=production
k8s.namespace.name=payments
host.name=worker-03
http.response.status_code=500
trace_id=abc123

That structure allows an AI system to correlate across services, clusters, nodes, requests, and deployments.


2. Semantic Conventions

Raw telemetry is not enough. The metadata needs consistent meaning.

OpenTelemetry Semantic Conventions define common names and attributes for operations and data across traces, metrics, logs, profiles, and resources.

This is crucial for AIOps because AI needs to compare like with like.

Without semantic conventions, one team may emit:

app = checkout

another may emit:

service = checkout-api

and another:

component = payments-checkout

An AIOps system then has to guess whether these are the same thing.

With standard conventions, the data becomes more machine-readable:

service.name = checkout-api
service.namespace = payments
deployment.environment = production
k8s.cluster.name = prod-eu-1

That makes correlation, incident grouping, ownership mapping, and root-cause analysis much stronger.


3. Multi-Signal Observability

Traditional monitoring was heavily metrics-focused.

Modern observability combines multiple signals:

Metrics  → What is happening numerically?
Logs → What discrete events occurred?
Traces → How did a request move through the system?
Profiles → Which code consumed CPU, memory, or wall time?
Events → What changed in the platform?

Kubernetes documentation still describes observability around metrics, logs, and traces as the main pillars for understanding cluster state, performance, and health. OpenTelemetry also describes observability signals as system outputs that describe application and platform activity.

For AIOps, this is fundamental.

A metric may say:

API latency is high.

A trace may say:

The latency is in the database query span.

A log may say:

Connection pool exhausted.

A deployment event may say:

New version deployed 5 minutes before the issue.

A profile may say:

CPU is being consumed by JSON serialisation in one function.

The AI can then produce a much better diagnosis than any single signal could provide.


4. Distributed Tracing and Context Propagation

Distributed tracing is one of the most important advances for AIOps.

In a monolith, a request might fail inside one process. In a microservices or cloud-native system, a single user request may cross:

Frontend
API gateway
Auth service
Checkout service
Payment service
Inventory service
Database
Message queue
External SaaS API

A trace connects those hops into one request journey.

For AIOps, tracing gives causal structure. It helps answer:

Where did the request slow down?
Which service returned the error?
Was the failure upstream or downstream?
Which tenant, region, node, or deployment was involved?

This makes root-cause analysis much more precise.

Without tracing, AIOps sees a pile of logs and metrics.

With tracing, it sees a connected execution path.


5. Exemplars: Linking Metrics to Traces

Another important advance is the ability to connect aggregate metrics to specific trace examples.

For example, a dashboard may show:

p99 latency = 2.4 seconds

But the engineer or AI needs to know:

Which actual request was slow?
What did that request do?
Which span caused the delay?

OpenTelemetry metrics support exemplars containing trace and span association fields, and Prometheus/OpenMetrics interoperability includes exemplar conversion rules.

For AIOps, exemplars are powerful because they bridge:

Metric anomaly → actual trace → logs from same request → root cause

That reduces guesswork.


6. Native Histograms and Better Latency Data

AIOps needs good latency distribution data, not just averages.

Average latency hides problems.

Example:

Average latency: 120 ms

That sounds fine, but the distribution may be:

95% of requests: 80 ms
4% of requests: 400 ms
1% of requests: 8 seconds

The 1% tail may be where real user pain exists.

Prometheus native histograms improve how latency and distribution data can be represented, and Prometheus native histograms with standard schemas can map to OpenTelemetry exponential histograms.

For AIOps, this improves:

anomaly detection
SLO burn analysis
performance regression detection
capacity planning
tail-latency investigation

AI needs distribution-aware telemetry to avoid drawing conclusions from misleading averages.


7. Telemetry Pipelines and Data Processing

Another major advance is the rise of programmable telemetry pipelines.

The OpenTelemetry Collector can receive, process, and export telemetry, and its processors can transform, filter, and enrich telemetry as it flows through a pipeline. Grafana Alloy also provides pipelines for telemetry signals such as Prometheus and OpenTelemetry, with support for logs, metrics, traces, and profiles.

This is vital for AIOps because raw telemetry is often messy.

You need to:

drop noisy fields
redact secrets
normalise labels
add environment metadata
add ownership information
route critical data differently
sample high-volume traces
preserve error traces
enrich logs with Kubernetes metadata
convert vendor-specific formats

For AIOps, the telemetry pipeline becomes the data preparation layer.

Bad pipeline:

AI receives noisy, inconsistent, high-volume telemetry.

Good pipeline:

AI receives enriched, normalised, relevant operational evidence.

That is the difference between useful AIOps and expensive confusion.


8. Continuous Profiling

Continuous profiling is another big step forward.

Metrics tell you that CPU is high.

Profiles tell you which code path is consuming CPU.

OpenTelemetry describes profiles as answering which code is responsible for consuming resources, complementing logs, metrics, and traces. The OpenTelemetry Profiles specification describes profiles as an emerging fourth observability signal alongside logs, metrics, and traces. Grafana Pyroscope describes continuous profiling as a systematic method for collecting and analysing performance data from production systems.

For AIOps, profiling helps move from:

The service is slow.

to:

The service is slow because 63% of CPU time is spent in JSON serialisation
inside checkout-api after the latest release.

That is much closer to actionable root cause.


9. eBPF-Based Observability

eBPF has significantly improved infrastructure and network observability.

Cilium describes itself as an eBPF-based solution for networking, observability, and security, providing visibility into workload connectivity. Hubble, built on Cilium, uses eBPF to provide dynamic visibility with detailed insight where needed.

For AIOps, eBPF is valuable because it can observe behaviour at the kernel and network layer without requiring every application to be perfectly instrumented.

It can help answer:

Which pod connected to which service?
Where are packets being dropped?
Is DNS failing?
Is the issue L3, L4, or L7?
Is network policy blocking traffic?
Is the service reachable?
Which process opened this connection?

This is especially important for Kubernetes, OpenStack, service mesh, GPU clusters, and distributed storage platforms.

For your type of environment, eBPF observability is particularly relevant because many failures happen below the application layer:

Neutron networking
Kubernetes CNI
DNS
load balancing
firewalling
pod-to-pod connectivity
GPU node networking
Ceph traffic
Slurm controller-to-worker communication

10. Topology-Aware Observability

AIOps cannot do strong root-cause analysis if it does not understand relationships.

It needs topology.

Example:

frontend
depends on checkout-api
depends on postgres
runs on k8s-worker-03
uses ceph-volume-17
backed by osd-4
runs on storage-node-02

This lets AIOps understand blast radius.

Without topology, AI sees isolated symptoms:

frontend errors
checkout latency
database timeout
Ceph OSD warning
node disk latency

With topology, it can infer:

Ceph OSD degradation on storage-node-02 is affecting postgres,
which is causing checkout-api latency and frontend errors.

This is one of the areas where observability has had to evolve from charts into graph-based operational context.


11. Change-Aware Observability

Most incidents are caused by change.

AIOps becomes much more effective when observability includes change events:

deployments
config changes
Terraform applies
Ansible runs
Kubernetes rollouts
OpenStack reconfigures
package upgrades
certificate renewals
DNS changes
firewall changes
feature flags
autoscaling events

AIOps needs to answer:

What changed before the incident?
Who changed it?
Was it automated?
Which services were affected?
Has this change caused problems before?

This is where GitLab, GitHub Actions, ArgoCD, Terraform, Ansible, Kubernetes events, OpenStack events, and audit logs become part of observability.

For example:

10:01 GitLab deployed checkout-api v2.4.1
10:03 pods restarted
10:04 p99 latency increased
10:05 HTTP 500s increased
10:06 SLO burn alert fired

The likely root cause is not “high latency.”

The likely root cause is the deployment.


12. SLO-Based Observability

Another advance is the move from infrastructure-centric alerts to service-centric SLOs.

Old alerting:

CPU > 90%
Disk > 80%
Pod restarted
Node memory high

Better alerting:

Checkout API availability below SLO
Payment latency budget burning too fast
Login error rate above user-impact threshold

For AIOps, SLOs provide priority.

Not every anomaly matters equally.

A CPU spike on a batch node may be fine.

A small increase in payment failures may be urgent.

SLO-based observability helps AIOps rank incidents by user impact rather than raw technical noise.


13. High-Cardinality and Dimensional Telemetry

Modern systems need dimensional analysis.

You need to slice by:

service
namespace
cluster
region
tenant
customer
version
endpoint
pod
node
GPU model
availability zone
database shard
queue
deployment

Prometheus uses a dimensional data model where time series are identified by a metric name and key-value labels, and PromQL allows teams to query, correlate, and transform time-series data.

For AIOps, dimensions are essential.

Instead of:

API latency is high.

you want:

API latency is high only for:
service=checkout-api
version=v2.4.1
region=eu-west
tenant=customer-a
endpoint=/payment/confirm

That turns a vague incident into a narrowed investigation.


14. Better Log Structure

AIOps performs much better with structured logs.

Bad log:

Something went wrong while processing request.

Better log:

{
"level": "error",
"service.name": "checkout-api",
"trace_id": "abc123",
"user_impact": true,
"order_id": "redacted",
"error.type": "DatabaseConnectionTimeout",
"db.system": "postgresql",
"k8s.pod.name": "checkout-api-7c9fd",
"deployment.version": "v2.4.1"
}

Structured logs let the AI search, group, and correlate events reliably.

For AIOps, this is a major difference.

Unstructured logs require interpretation.

Structured logs provide evidence.


15. Observability for Automation

AIOps is not only about diagnosis. It also needs verification.

Before remediation:

Is the service unhealthy?
What is the likely root cause?
Is the proposed action safe?

After remediation:

Did the error rate fall?
Did latency recover?
Did the pod restart cleanly?
Did the SLO burn rate stabilise?
Did the same alert return?

Observability provides the feedback loop for automation.

Without observability, automation is blind.

A mature AIOps loop looks like this:

Detect

Correlate

Diagnose

Recommend

Approve

Execute

Verify

Learn

The “verify” and “learn” stages depend heavily on observability.


16. AI-Readable Operational Context

The latest practical advance is making observability data usable by AI agents.

That means exposing operational systems through APIs, query layers, or protocols such as MCP-style tool access.

The AI needs controlled access to:

metrics queries
log search
trace lookup
profile analysis
Kubernetes state
OpenStack state
Slurm queue state
Ceph health
GitLab deployments
Ansible runbooks
incident history
service ownership

This turns observability from something humans look at into something AI can query and reason over.

For example:

User asks:
"Why did gpu-test-01 fail to start?"

AI queries:
OpenStack state
Nova logs
Neutron events
Prometheus GPU metrics
Slurm state
Kubernetes GPU operator status
Ceph health
recent config changes

AI replies:
"Nova failed to schedule the VM because the requested PCI alias is not defined.
Neutron and storage are healthy. The failure is isolated to Nova PCI configuration."

That is observability becoming operational intelligence.


How These Advances Make AIOps Effective

The relationship is simple:

Observability advanceWhat it gives AIOps
OpenTelemetryStandard telemetry collection
Semantic conventionsConsistent metadata
MetricsQuantitative system health
LogsEvent-level explanation
TracesRequest-level causality
ProfilesCode-level resource attribution
ExemplarsLink from metric anomaly to trace
eBPFKernel/network visibility
TopologyDependency and blast-radius context
Change events“What changed?” analysis
SLOsBusiness/user-impact priority
Telemetry pipelinesClean, enriched, governed data
Structured logsMachine-readable evidence
Automation feedbackSafe remediation verification

For Your OpenStack / Kubernetes / Slurm / GPU Homelab

For your environment, the observability stack that would make AIOps effective should include:

Metrics:
Prometheus / Mimir

Logs:
Loki

Traces:
Tempo

Profiles:
Pyroscope

Collection and pipelines:
OpenTelemetry Collector or Grafana Alloy

Dashboards and exploration:
Grafana

Alerting:
Alertmanager

Kubernetes network observability:
Cilium / Hubble / eBPF

Change context:
GitLab CI/CD, Ansible logs, OpenStack reconfigure events

Infrastructure state:
OpenStack, Nova, Neutron, Glance, Cinder, Ceph

GPU/HPC state:
Slurm, NVIDIA DCGM exporter, nvidia-smi, GPU Operator if using Kubernetes

Automation:
Ansible playbooks with human approval

AI access:
MCP-style tool interface to query metrics, logs, traces, infrastructure, and runbooks

The important design principle is:

Do not just collect telemetry.
Make telemetry correlated, structured, enriched, searchable, and safe for AI to use.

That is what turns observability into a real AIOps foundation.

AI Observability

AI is not killing observability as a discipline, but it is fundamentally changing how observability is done. The traditional model of “collect everything, store everything, and let humans investigate later” is becoming increasingly impractical in AI-driven infrastructures.

1. Telemetry volume is exploding

Modern systems produce far more telemetry than they did five years ago.

An AI factory may contain:

  • Tens of thousands of GPUs
  • Hundreds of thousands of CPU cores
  • High-speed fabrics (RoCE, InfiniBand)
  • Kubernetes
  • Distributed storage (Ceph, Lustre, GPFS)
  • AI inference services
  • LLM gateways

Each component exports metrics, logs, traces and events.

For example:

2020

100 servers

100 million metrics/day

2026

20,000 GPUs
30,000 CPUs
5,000 switches



Several trillion data points/day

Humans cannot meaningfully explore that volume.


2. Dashboards don’t scale

Traditional observability assumes people sit looking at Grafana dashboards.

Reality:

  • nobody watches 400 dashboards
  • nobody remembers 2,000 PromQL queries
  • nobody notices slow drift

Instead people increasingly ask:

“Why did training become slower?”

AI investigates.

Not humans.


3. Alert fatigue becomes impossible

Large organisations often generate

  • 50,000 alerts/day
  • 100,000 log anomalies/day

Historically:

Prometheus
↓
Alertmanager
↓
PagerDuty
↓
Human

Future:

Prometheus
↓
AI correlation
↓
Root cause
↓
Human receives one explanation

Instead of:

127 alerts

Engineer receives

GPU node gpu-128 experienced ECC errors causing NCCL retries which slowed training by 18%.


4. Humans don’t query telemetry anymore

Traditional workflow

Grafana
↓
Zoom
↓
PromQL
↓
Logs
↓
Tempo
↓
Find issue

Future

"Why are customer requests slower?"
↓
AI
↓
queries everything
↓
returns explanation

Natural language replaces much of manual exploration.


5. AI is becoming the first investigator

Large enterprises increasingly build systems like:

Telemetry
↓
LLM
↓
Reasoning
↓
Correlation
↓
Recommendation

Instead of asking engineers to join the dots.


6. Sampling changes everything

Historically:

Store every log.

Now:

AI decides

Keep
Discard
Summarise
Compress

Observability becomes intelligent instead of passive.


7. Root cause becomes graph reasoning

Today’s tools often correlate:

metric
+
trace
+
log

Future systems correlate:

  • topology
  • Kubernetes
  • network
  • storage
  • deployments
  • Git commits
  • feature flags
  • incidents
  • Slack discussions
  • runbooks

into one knowledge graph.

AI reasons across all of it.


8. AI reduces the need for experts

Today an SRE may spend years learning:

  • PromQL
  • LogQL
  • TraceQL
  • Elasticsearch
  • Loki
  • Tempo
  • Mimir
  • Kubernetes
  • networking

Future engineers may simply ask:

Why did latency increase?

The AI handles the underlying queries.

Expertise shifts from query syntax to validating conclusions and understanding system design.


9. Observability vendors are changing

Every major platform is investing heavily in AI assistants.

Examples include:

  • Grafana Labs
  • Datadog
  • Dynatrace
  • New Relic
  • Elastic
  • Splunk

They’re moving from:

  • dashboards

to:

  • AI copilots
  • automatic investigations
  • causal analysis
  • anomaly explanations
  • remediation suggestions

10. Cost is becoming the biggest problem

Storage costs are growing faster than engineering budgets.

For many enterprises:

Telemetry generated

100 TB/day

Engineers never inspect 99% of it.

AI can:

  • summarize repetitive logs
  • identify low-value telemetry
  • adapt sampling dynamically
  • retain only diagnostically useful data

This can significantly reduce storage and processing costs while preserving investigative value.


What is actually dying?

It is not observability itself, but the manual workflows around it.

Declining:

  • Manual dashboard creation
  • Hand-written alert rules for every scenario
  • Endless log searching
  • Human correlation across tools
  • Memorizing complex query languages

Growing:

  • AI-assisted investigations
  • Automated root-cause analysis
  • Predictive anomaly detection
  • Natural-language querying
  • Knowledge-graph reasoning
  • Automated remediation

What this means for SREs

For someone with your background in observability platforms, this shift is an opportunity rather than a threat. The value moves away from being the person who can write the most intricate PromQL query and toward designing the telemetry architecture, ensuring data quality, integrating AI agents with monitoring systems, and deciding when AI-generated conclusions can be trusted.

The emerging role is less “dashboard builder” and more Observability Platform Engineer or AI Operations (AIOps) Engineer—someone who builds systems that enable AI to investigate production issues safely and accurately. That combination of platform engineering, distributed systems, and AI integration is likely to be in increasing demand over the next decade.

What is AI Observability?

AI-era observability is moving from human-driven inspection to machine-assisted reasoning over telemetry, topology, history and operational knowledge.

The key shift is this:

Old observability:

Metrics + logs + traces

Dashboards and alerts

Human investigates

Human decides

Human fixes


AI-era observability:

Metrics + logs + traces + topology + deployments + runbooks + incidents

AI correlation and reasoning layer

Probable cause, blast radius, next action

Human approval or automated remediation

Below is a detailed breakdown of the six areas.


1. AI-assisted investigations

What it means

AI-assisted investigation is where an AI system acts like a junior SRE investigator sitting beside you.

It does not necessarily fix the issue automatically. Its main job is to reduce the time spent asking basic investigative questions.

Instead of you manually jumping between:

Grafana → Prometheus/Mimir → Loki → Tempo → Kubernetes → Git → Slack → Runbooks

you ask something like:

Why did checkout latency increase after 14:05?

The AI then queries multiple systems and returns a structured investigation.


What it does

A good AI investigation assistant can:

  • Detect the relevant service, namespace, cluster or tenant.
  • Pull related metrics.
  • Search logs around the incident window.
  • Inspect traces for slow spans.
  • Compare current behaviour against baseline behaviour.
  • Check recent deployments.
  • Check Kubernetes events.
  • Check node, pod, container and network health.
  • Retrieve relevant runbooks.
  • Summarise likely causes.
  • Recommend next diagnostic steps.

Example

You ask:

Why is the inference API slower?

The AI investigates:

1. Latency increased at 10:17.
2. p95 rose from 420 ms to 1.8 s.
3. Error rate did not increase.
4. GPU utilisation remained high.
5. Queue depth increased.
6. New model version was deployed at 10:12.
7. Logs show repeated batching timeout warnings.
8. Traces show delay before GPU execution, not during execution.

Result:

Likely issue:
The model service is queueing requests before GPU execution.

Probable cause:
The new batching configuration increased max_batch_wait_ms from 20 ms to 250 ms.

Recommended action:
Rollback batching config or reduce batch wait threshold.

That is much faster than manually checking ten dashboards.


What data it needs

AI-assisted investigation works best when it has access to:

Metrics:
- RED metrics: rate, errors, duration
- USE metrics: utilisation, saturation, errors
- Kubernetes pod/node metrics
- GPU metrics
- Network metrics
- Storage metrics

Logs:
- Application logs
- Kubernetes events
- System logs
- Ingress/controller logs
- Deployment logs

Traces:
- Request path
- Slow spans
- Upstream/downstream dependencies
- Database/storage/API calls

Context:
- Deployment history
- Git commits
- Feature flags
- Config changes
- Runbooks
- Incident history
- Service ownership

Without context, AI just summarises telemetry. With context, it can investigate.


SRE value

For SREs, this means less time doing mechanical investigation and more time validating the diagnosis.

The future SRE skill is not just:

Can I write PromQL?

It becomes:

Can I design telemetry so AI can reason correctly?
Can I validate the AI's conclusion?
Can I prevent unsafe remediation?
Can I encode good operational knowledge into the platform?

2. Automated root-cause analysis

What it means

Automated root-cause analysis, or automated RCA, is the process of identifying the most likely initiating cause of a production issue without relying entirely on manual human correlation.

It tries to answer:

What actually started the incident?

Not merely:

What symptoms are currently visible?

This distinction matters.


Symptom versus root cause

Example incident:

Customer latency is high.
API pods are slow.
Database queries are slow.
Storage latency is high.
Ceph OSDs are rebalancing.
One storage node has a failing disk.

The symptoms are:

High API latency
Slow database responses
Increased request duration
More timeout warnings

The probable root cause is:

A failing disk caused Ceph recovery/rebalancing,
which increased storage latency,
which slowed the database,
which slowed the API.

Automated RCA attempts to build that causal chain.


How automated RCA works

There are several techniques.

1. Temporal correlation

The system checks what changed first.

10:01 disk errors begin
10:03 Ceph recovery starts
10:05 storage latency rises
10:07 database latency rises
10:09 API latency rises
10:10 customer alerts fire

The earliest credible abnormal event is often close to the root cause.


2. Topology-aware analysis

The system understands dependencies.

frontend

checkout-api

postgres

ceph/rbd

osd-node-07

If osd-node-07 is unhealthy, and all dependent systems are degraded, the RCA engine can infer blast radius.


3. Change correlation

The system checks recent changes:

Deployments
Config changes
Feature flags
Kernel updates
Node drains
Network changes
Storage migrations
Certificate rotations
DNS changes
Autoscaling events

Many incidents are change-induced. A useful RCA system always asks:

What changed recently?

4. Statistical anomaly ranking

The system ranks abnormal signals.

For example:

Signal                         Abnormality score
GPU ECC errors 0.98
NCCL retry count 0.94
Training step duration 0.91
CPU usage 0.22
Memory usage 0.18

The AI focuses on the strongest abnormal signals.


5. Causal graph reasoning

This is more advanced.

Instead of treating metrics as isolated time series, the system builds a causal model:

Bad disk
→ Ceph recovery
→ Storage latency
→ Database latency
→ API latency
→ Customer impact

This is much closer to how an experienced SRE thinks.


Example automated RCA output

Incident:
Checkout latency p95 increased from 300 ms to 2.4 s.

Likely root cause:
PostgreSQL read latency increased due to degraded Ceph RBD volume performance.

Evidence:
- API latency increased at 13:42.
- PostgreSQL read latency increased at 13:39.
- Ceph pool latency increased at 13:36.
- OSD 12 reported slow ops and disk errors at 13:34.
- No relevant application deployment occurred in the previous hour.

Blast radius:
- checkout-api
- payment-api
- order-history-api

Recommended action:
- Mark OSD 12 out if disk errors continue.
- Move affected workload if possible.
- Check Ceph recovery/backfill limits.
- Consider temporarily scaling API timeout thresholds.

What makes automated RCA hard

Automated RCA is difficult because distributed systems are messy.

Common problems:

Correlation is not causation.
Multiple things can break at once.
Telemetry may be missing.
Logs may be noisy.
Clocks may not be perfectly synchronised.
Service dependency maps may be stale.
The root cause may be outside the monitored system.

This is why good automated RCA usually gives:

Probable cause
Confidence level
Supporting evidence
Contradicting evidence
Recommended next checks

It should not pretend to be certain when it is not.


3. Predictive anomaly detection

What it means

Predictive anomaly detection tries to detect abnormal behaviour before it becomes a major incident.

Traditional alerting says:

Alert when disk usage > 90%.

Predictive alerting says:

Disk usage is growing at a rate that will hit 90% in 11 hours.

That is a major shift.


Traditional threshold alerting

Example:

Alert: DiskAlmostFull
Condition: disk_used_percent > 90

This is simple and useful, but it misses context.

A disk at 85% may be fine if it grows slowly.

A disk at 60% may be dangerous if it is growing rapidly.


Predictive anomaly detection

Predictive systems look at behaviour over time:

Normal pattern:
- CPU rises during business hours
- drops overnight
- spikes during batch processing

Abnormal pattern:
- CPU rises at midnight
- no scheduled job exists
- memory grows continuously
- request rate is normal

The system detects that the pattern is unusual, even if no hard threshold has been crossed.


Types of predictive anomalies

1. Trend-based prediction

Useful for capacity planning.

Disk usage will reach 90% in 3 days.
Mimir object storage will exceed budget in 12 days.
Kafka partition disk will fill in 9 hours.
Ceph pool will hit near-full ratio this weekend.

2. Seasonality-aware anomaly detection

Useful for normal daily/weekly cycles.

Example:

CPU at 80% at 10:00 Monday may be normal.
CPU at 80% at 03:00 Sunday may be abnormal.

The system learns expected patterns.


3. Multivariate anomaly detection

Looks at several signals together.

For example:

Request rate: normal
Error rate: normal
Latency: high
CPU: normal
Database latency: high
Network retransmits: high

Individually, some metrics may not trigger alerts. Together, they reveal an abnormal condition.


4. Behavioural drift detection

Useful in AI and ML platforms.

Example:

Training jobs are completing successfully,
but average step time has increased by 12% over two weeks.

No incident has occurred yet, but performance is drifting.


5. Saturation prediction

Very useful for SRE.

GPU memory saturation likely within 40 minutes.
Kubernetes node memory pressure likely in 2 hours.
Ceph recovery will saturate backend network.
Kafka consumer lag will exceed SLO in 25 minutes.

Example

A predictive anomaly detector observes:

Mimir ingest rate: stable
Object storage write latency: slowly increasing
Compactor duration: increasing
Query latency: increasing
Store-gateway cache hit rate: decreasing

It predicts:

Within 6 hours, users will experience slow dashboard loads.

The remediation might be:

Scale store-gateways.
Check object storage latency.
Increase cache.
Review compactor backlog.

Why it matters

Predictive anomaly detection changes operations from:

React after customer impact

to:

Intervene before customer impact

That is the core SRE value.


4. Natural-language querying

What it means

Natural-language querying allows engineers to ask operational questions in plain English instead of writing PromQL, LogQL, TraceQL, SQL or Elasticsearch queries manually.

Example:

Show me p95 latency for checkout-api over the last 6 hours,
split by Kubernetes namespace.

The AI converts that into the right query.


Traditional workflow

You need to know the query language:

histogram_quantile(
0.95,
sum by (le, namespace) (
rate(http_request_duration_seconds_bucket{
service="checkout-api"
}[5m])
)
)

With natural-language querying:

What is checkout-api p95 latency by namespace for the last 6 hours?

The AI generates and executes the query.


Where this is useful

Natural-language querying is useful across:

Metrics:
- Prometheus
- Mimir
- Thanos
- VictoriaMetrics

Logs:
- Loki
- Elasticsearch/OpenSearch
- ClickHouse

Traces:
- Tempo
- Jaeger
- OpenTelemetry backends

Databases:
- PostgreSQL
- BigQuery
- Snowflake
- ClickHouse

Cloud APIs:
- Kubernetes
- AWS
- Azure
- GCP
- OpenStack

Example questions

Which services had the largest increase in error rate in the last hour?

Show me pods that restarted after the latest deployment.

Find logs for payment-api where timeout errors increased.

Which traces spent the most time waiting on PostgreSQL?

Which Kubernetes nodes have high network retransmits?

Show me Ceph OSDs with rising latency and degraded placement groups.

Which GPU nodes show ECC errors or thermal throttling?

The important part: semantic mapping

Natural-language querying is not just text-to-query.

It needs to understand your telemetry naming.

For example, you may ask:

Show API latency.

But your metrics may be called:

http_request_duration_seconds_bucket
nginx_ingress_controller_request_duration_seconds_bucket
istio_request_duration_milliseconds_bucket
app_http_server_duration_bucket

The AI needs a semantic layer that maps human concepts to real telemetry.


Good natural-language querying needs

Metric catalogue
Label documentation
Service ownership map
Namespace conventions
Dashboard metadata
Runbook links
Known-good query examples
SLO definitions
Deployment metadata

Without that, the AI may generate syntactically valid but operationally useless queries.


Risk: hallucinated queries

Natural-language querying can be dangerous if it invents metric names.

Bad output:

rate(checkout_latency_seconds[5m])

But that metric may not exist.

Better behaviour:

I could not find a metric named checkout_latency_seconds.
I found http_request_duration_seconds_bucket with service="checkout-api".
Using that instead.

The AI should verify queries against the actual telemetry backend.


SRE impact

SREs will still need to understand PromQL, LogQL and traces, but less time will be spent manually composing queries.

The valuable skill becomes designing the semantic layer:

Good metric names
Useful labels
Consistent service metadata
Accurate ownership data
Clear runbooks
Well-documented SLOs

5. Knowledge-graph reasoning

What it means

Knowledge-graph reasoning connects operational facts into a graph so AI can reason over relationships.

Traditional observability stores data like this:

Metric:
checkout-api p95 latency = 2.1s

Log:
timeout connecting to postgres

Trace:
checkout-api → postgres took 1.8s

Kubernetes:
postgres pod moved to node-12

Infrastructure:
node-12 has disk pressure

A knowledge graph connects those facts:

checkout-api
depends_on → postgres
runs_in → namespace prod
owned_by → payments-team

postgres
runs_on → node-12
uses → ceph-rbd-volume-44

node-12
has_condition → disk_pressure

ceph-rbd-volume-44
backed_by → ceph-pool-prod

Now the AI can reason across relationships.


Why graphs matter

Most incidents are not isolated.

They involve chains:

Application → runtime → Kubernetes → node → network → storage → hardware

Dashboards show symptoms. Graphs show relationships.


Example graph

Customer impact

checkout-api latency

postgres query latency

RBD volume latency

Ceph OSD slow ops

failing NVMe device

A graph-based system can move up and down this chain.


What goes into the graph

A strong observability knowledge graph includes:

Services
APIs
Databases
Queues
Kubernetes namespaces
Pods
Nodes
Clusters
Storage volumes
Ceph pools
Network devices
Load balancers
Ingress controllers
Deployments
Git commits
Feature flags
SLOs
Alerts
Incidents
Runbooks
Owners
Escalation paths
Cloud resources
OpenStack projects
GPU nodes
Training jobs

How the graph is built

Data sources may include:

Kubernetes API
Prometheus/Mimir labels
OpenTelemetry resource attributes
Service mesh telemetry
CMDB
Terraform state
GitOps repositories
CI/CD systems
Incident management tools
Cloud APIs
OpenStack APIs
Ceph APIs
Network controllers
Runbooks and docs

The graph is continuously updated.


Example reasoning

Question:

Why are training jobs slower on rack 3?

The graph helps the AI discover:

Training-job-982
runs_on → gpu-node-31, gpu-node-32, gpu-node-33
located_in → rack-3
uses_network → leaf-switch-3a
uses_storage → lustre-client
depends_on → metadata-server-2

Telemetry shows:

leaf-switch-3a has rising packet drops
NCCL retries increased
GPU utilisation has sawtooth pattern
training step time increased

AI conclusion:

The training slowdown is likely caused by network instability on rack 3,
not by GPU compute saturation.

That is knowledge-graph reasoning.


Why this matters for AI data centres

AI/HPC environments are dependency-heavy.

A single training workload may depend on:

GPU health
GPU memory
NVLink/NVSwitch
PCIe
RoCE/InfiniBand
Leaf-spine network
Storage bandwidth
Metadata servers
Container runtime
Kubernetes scheduler
Slurm scheduler
Image registry
Secrets
DNS
Authentication
Object storage

A flat dashboard cannot represent that well. A graph can.


6. Automated remediation

What it means

Automated remediation is when the system not only detects and diagnoses an issue, but also takes corrective action.

This is the most powerful and most dangerous part of AI-era observability.

It moves from:

Observe → Alert → Human fixes

to:

Observe → Diagnose → Decide → Act → Verify

Simple automated remediation

Low-risk examples:

Restart a failed pod.
Scale a deployment from 3 to 5 replicas.
Clear a stuck job.
Rotate a saturated log file.
Drain a bad Kubernetes node.
Open an incident ticket.
Create a Slack/PagerDuty summary.
Rollback a known-bad deployment.
Increase queue consumers.

Advanced automated remediation

Higher-risk examples:

Move workloads away from degraded storage.
Change Ceph recovery/backfill settings.
Disable a feature flag.
Rebalance Kafka partitions.
Quarantine a GPU node.
Remove a bad node from a load balancer.
Apply a network policy change.
Trigger disaster recovery failover.
Patch a vulnerable service.

These require stronger guardrails.


The remediation loop

A safe remediation system should work like this:

1. Detect
Something abnormal happened.

2. Diagnose
Determine probable cause and confidence.

3. Propose
Generate a remediation plan.

4. Check policy
Is this action allowed?
Is the blast radius acceptable?
Is approval required?

5. Act
Execute the change.

6. Verify
Did the metric improve?
Did errors reduce?
Did customer impact stop?

7. Roll back
If not improved, revert or escalate.

8. Learn
Record the incident and outcome.

Example

Issue:

checkout-api error rate increased after deployment.

AI investigation:

New version deployed at 09:03.
Errors began at 09:05.
Only pods running version v2.7.4 are affected.
Previous version v2.7.3 had no errors.

Remediation proposal:

Rollback checkout-api from v2.7.4 to v2.7.3.

Policy check:

Allowed because:
- service has rollback automation
- error rate exceeds SLO threshold
- last known-good version exists
- no database migration detected

Action:

kubectl rollout undo deployment/checkout-api

Verification:

Error rate returned to baseline after 4 minutes.
p95 latency returned to normal.
Incident summary created.

Guardrails are essential

Automated remediation must not be a reckless agent with production write access.

Good guardrails include:

Read-only by default
Approval required for high-risk actions
Change windows
Blast-radius limits
Dry-run mode
Policy-as-code
RBAC
Audit logs
Rollback plans
Rate limits
Canary execution
Human confirmation for destructive actions

For example:

Allowed automatically:
- restart one unhealthy pod
- scale a stateless service within limits
- create an incident ticket

Requires approval:
- drain production node
- rollback payment service
- modify firewall/network policy
- change Ceph recovery settings
- fail over database

How these six areas fit together

They are not separate ideas. They form a pipeline.

Natural-language querying

Lets humans ask better questions

AI-assisted investigations

Gathers evidence automatically

Knowledge-graph reasoning

Understands relationships and dependencies

Automated root-cause analysis

Identifies probable initiating cause

Predictive anomaly detection

Finds issues before they become incidents

Automated remediation

Fixes or mitigates the issue

A mature AI observability platform combines all six.


Practical architecture for an AI observability platform

A realistic architecture could look like this:

Telemetry sources
├─ Prometheus / Mimir metrics
├─ Loki logs
├─ Tempo traces
├─ Kubernetes events
├─ Ceph / storage metrics
├─ GPU metrics
├─ Network telemetry
├─ CI/CD events
├─ Git commits
└─ Incident history



Data normalization layer
├─ OpenTelemetry attributes
├─ Service naming standards
├─ Environment labels
├─ Owner/team labels
└─ SLO metadata



Context layer
├─ Runbooks
├─ Architecture docs
├─ Past incidents
├─ Known failure modes
├─ Deployment history
└─ Dependency maps



AI reasoning layer
├─ LLM
├─ RAG over runbooks/docs
├─ Query generation
├─ Anomaly detection
├─ Causal graph reasoning
└─ RCA ranking



Action layer
├─ Human-readable incident summary
├─ Suggested next steps
├─ Ticket creation
├─ Slack/PagerDuty update
├─ Safe automation
└─ Approved remediation

What you would build first as an SRE

I would not start with fully automated remediation. That is too risky.

The sensible maturity path is:

Stage 1: AI-assisted read-only investigation

Build a tool that can answer:

What changed?
What alerts fired?
What services are affected?
What logs are unusual?
What traces are slow?
What runbook applies?

No write actions.


Stage 2: Natural-language query assistant

Allow engineers to ask:

Show me p95 latency by service.
Find logs for this incident window.
Show me failed pods after the deployment.
Compare today’s error rate with yesterday.

The assistant should show the generated query so the engineer can verify it.


Stage 3: Incident summariser

Generate structured summaries:

Incident:
Impact:
Start time:
Affected services:
Probable cause:
Evidence:
Actions taken:
Current status:
Recommended next steps:

This alone saves huge operational time.


Stage 4: RCA recommendation engine

Add correlation with:

Deployments
Kubernetes events
Node health
Storage health
Network telemetry
Recent config changes

Output probable root cause with confidence.


Stage 5: Predictive alerting

Start with safer predictions:

Disk will fill.
Object storage usage will exceed budget.
Kafka lag will breach SLO.
Ceph pool will hit near-full.
Certificate will expire.
GPU nodes are showing increasing ECC errors.

Stage 6: Human-approved remediation

The AI proposes actions, but humans approve.

Example:

Recommended action:
Drain node gpu-17 and reschedule workloads.

Reason:
GPU ECC errors increased and training retries are affecting jobs.

Approval required:
Yes.

Stage 7: Limited automatic remediation

Only allow automation for narrow, reversible, low-risk actions.

Restart crashed pod
Scale stateless deployment
Reopen failed consumer
Create incident ticket
Disable noisy alert temporarily with expiry

Main risks

AI observability can go wrong if the system has poor telemetry or too much authority.

1. Bad telemetry in, bad reasoning out

If labels are inconsistent, traces are incomplete, or logs are unstructured, AI conclusions will be weak.


2. Hallucinated root cause

The AI may sound confident while being wrong.

Always require:

Evidence
Confidence
Alternative theories
Query links
Raw data references

3. Unsafe remediation

A bad automated action can make an incident worse.

Example:

AI sees high memory.
AI restarts all pods.
All pods restart at once.
Outage gets worse.

That is why blast-radius control matters.


4. Hidden cost explosion

AI investigation can generate expensive backend queries.

A poorly controlled AI assistant may run huge queries across logs, traces and metrics.

You need:

Query limits
Timeouts
Caching
Sampling
Tenant controls
Cost visibility

5. Security and access control

The AI should not see or do everything.

It needs RBAC:

Read-only access for most users
Sensitive log masking
No secret exposure
Audit trail
Approval for write actions
Tenant isolation

The big picture

These six capabilities are the future of observability:

CapabilityMain purposeHuman role
AI-assisted investigationsSpeed up incident analysisValidate findings
Automated RCAIdentify probable causeJudge evidence
Predictive anomaly detectionPrevent incidents earlierTune models and thresholds
Natural-language queryingMake telemetry easier to accessVerify generated queries
Knowledge-graph reasoningUnderstand system relationshipsMaintain accurate topology
Automated remediationFix or mitigate issuesDefine guardrails and approve risk

The core change is this:

Observability is no longer just about collecting telemetry.

It is becoming a reasoning system over telemetry.

For SREs, the opportunity is to become the person who builds and governs that reasoning system: telemetry quality, context, automation safety, incident workflows, and trust boundaries.

Commercial AI Observability

Commercial companies are building AI into observability in two directions:

  1. AI for observability — using AI to investigate, correlate, explain, predict and remediate production issues.
  2. Observability for AI — monitoring LLMs, agents, RAG pipelines, vector databases, model quality, hallucinations, token cost, latency, drift and safety.

So the product shift is not just “add a chatbot to dashboards.” The bigger move is toward an AI operations layer that sits above metrics, logs, traces, events, topology and runbooks.

Telemetry + topology + deployments + logs + traces + incidents + runbooks

AI reasoning layer

Explain issue → find cause → predict risk → recommend/execute action

1. Datadog

Datadog is building AI into its platform around Bits AI, Watchdog, and LLM/Agent Observability.

Datadog’s Watchdog is its AI engine for automated alerts, insights and root-cause analysis across Datadog telemetry. It continuously monitors infrastructure and surfaces important signals to help teams detect, troubleshoot and resolve issues.

Datadog’s Bits AI SRE is positioned as an always-on AI SRE agent that helps handle troubleshooting and alerts, with Datadog describing it as able to pinpoint root causes faster by using Datadog’s incident and telemetry context.

Datadog is also pushing Bits AI Agents and Agent Builder, where the platform can build custom AI agents that investigate issues, make decisions and take action using Datadog and third-party data, with prebuilt actions across cloud, security, CI/CD and collaboration tooling.

For the second direction, Datadog has Agent Observability / LLM Observability, aimed at tracing, evaluating and improving LLM-powered applications and AI agents. Datadog says each LLM application request can be represented as a trace, allowing teams to investigate root cause, operational performance, quality, privacy and safety.

In plain SRE terms, Datadog is building:

Datadog AI direction:

Watchdog
→ automatic anomaly detection
→ automated insights
→ RCA suggestions

Bits AI
→ natural-language investigation
→ AI SRE assistant
→ incident summarisation
→ workflow automation

Bits AI Agents
→ custom agentic workflows
→ investigation agents
→ remediation/documentation agents

LLM / Agent Observability
→ traces for LLM calls
→ prompt/response monitoring
→ quality, privacy, safety checks
→ AI-agent debugging

Datadog is also doing deeper model work: its Toto time-series foundation model is specifically designed for observability time-series forecasting and was trained partly on Datadog observability data.


2. Dynatrace

Dynatrace has probably been the most explicit about putting causal AI at the centre of observability.

Its AI engine is Davis AI / Dynatrace Intelligence. Dynatrace describes its AI approach as combining predictive AI, causal AI and generative AI over unified observability and security data to automate workflows.

Dynatrace’s key differentiator is that it does not want the AI to merely correlate metrics. It wants the platform to understand causality: what caused what, what depends on what, and what failure actually triggered the incident. Dynatrace describes causal AI as using causal and deterministic techniques to determine underlying causes and effects rather than just relying on correlation.

Dynatrace also presents Dynatrace Intelligence as combining deterministic insights with agentic action for prevention, remediation and optimisation at scale.

For AI workloads, Dynatrace has AI and LLM Observability for monitoring, optimising and securing generative AI apps, LLMs and agentic workflows, with emphasis on performance, explainability and compliance.

In SRE terms, Dynatrace is building:

Dynatrace AI direction:

Davis AI / Dynatrace Intelligence
→ anomaly detection
→ causal root-cause analysis
→ topology-aware problem detection
→ predictive risk detection
→ generative explanations
→ workflow automation

Causal AI
→ dependency-aware analysis
→ fault-tree-style reasoning
→ root cause, not just symptom correlation

AI and LLM Observability
→ GenAI app monitoring
→ LLM and agentic workflow visibility
→ explainability
→ compliance-oriented monitoring

The important point: Dynatrace is trying to make observability less like “search through telemetry” and more like automated dependency-aware diagnosis.


3. Splunk

Splunk is building AI into observability through Splunk AI Assistant in Observability Cloud, broader AI Observability, and AI/agent monitoring.

Splunk’s AI Assistant in Observability Cloud uses observability data from metrics, traces, logs and alerts through a chat interface inside Splunk Observability Cloud.

Splunk says the AI Assistant can analyze data across APM, Infrastructure Monitoring, Database Monitoring, RUM and log analytics to help with root-cause analysis.

Splunk is also building “observability for AI” capabilities. Its Splunk Observability for AI is described as full-fidelity monitoring and troubleshooting across AI applications and the AI infrastructure components used to build them.

Splunk’s AI Agent Monitoring aims to correlate degraded AI agent/model performance and track operational metrics such as latency and errors alongside quality/security metrics such as hallucinations, bias, drift, accuracy, cost and token usage.

Splunk’s AI Observability positioning is broader: observe and optimise performance, quality, cost and security across agents, LLMs, vector databases and infrastructure.

In SRE terms, Splunk is building:

Splunk AI direction:

AI Assistant in Observability Cloud
→ natural-language investigations
→ logs + metrics + traces + alerts analysis
→ RCA assistance
→ incident summarisation

AI Observability
→ AI application monitoring
→ AI infrastructure monitoring
→ agent performance tracking
→ LLM quality and safety monitoring

AI Agent Monitoring
→ latency and errors
→ hallucination tracking
→ bias/drift/accuracy
→ token and cost visibility
→ model and agent reliability

Splunk’s direction is very aligned with its historical strength: search, correlation and operational analytics, now wrapped in AI-assisted investigation and AI workload monitoring.


4. New Relic

New Relic is building AI into its platform through New Relic AI, AI-powered observability features, and AI Monitoring / LLM observability.

New Relic says New Relic AI can help instrument systems, generate system health reports and identify alert coverage gaps for full-stack observability.

New Relic has also positioned its platform as AI-powered observability that correlates telemetry across the stack to isolate root cause and reduce operational toil.

For LLM applications, New Relic AI monitoring captures telemetry from AI-powered apps through APM agents and collects data from external LLMs and vector stores.

New Relic’s AI monitoring focuses on troubleshooting, comparing and optimising LLM prompts and responses for performance, cost and quality issues such as hallucination, bias and toxicity.

It also supports LLM observability through OpenLIT integration, which automatically generates traces and metrics for LLM and VectorDB performance and cost analysis.

In SRE terms, New Relic is building:

New Relic AI direction:

New Relic AI
→ AI assistant for DevOps
→ system health reports
→ alert coverage analysis
→ instrumentation help

AI-powered observability
→ telemetry correlation
→ root-cause isolation
→ faster troubleshooting

AI Monitoring / LLM Observability
→ prompt/response analysis
→ LLM latency and error tracking
→ cost analysis
→ hallucination, bias and toxicity signals
→ VectorDB visibility

New Relic’s direction is about making its “all-in-one observability” platform more assistant-driven and making AI workloads first-class observable systems.


What they are all converging on

All four vendors are converging on the same broad architecture:

1. Collect telemetry
metrics, logs, traces, events, profiles, topology

2. Normalize context
services, owners, deployments, dependencies, SLOs, runbooks

3. Apply AI
anomaly detection, query generation, summarisation, RCA, prediction

4. Explain
what happened, why it happened, what changed, what is affected

5. Act
create ticket, page team, suggest fix, trigger workflow, remediate safely

6. Observe AI itself
LLM calls, prompts, responses, token cost, model quality, hallucinations,
safety, drift, vector DBs, RAG pipelines, agent workflows

The big product categories are:

AI capabilityWhat vendors are building
AI assistantChat interface over observability data
AI SRE agentInvestigates incidents and proposes actions
Automated RCAFinds likely root cause using telemetry and topology
Predictive anomaly detectionSpots problems before thresholds are breached
Natural-language queryingConverts plain English into PromQL, LogQL, SQL, trace/log queries
Incident summarisationExplains impact, timeline, evidence and next steps
Runbook automationRecommends or triggers operational workflows
AI workload monitoringMonitors LLMs, agents, prompts, responses, cost and quality
Governance/safetyTracks hallucination, toxicity, bias, privacy and compliance risks
Cost optimisationReduces telemetry waste and tracks LLM/token spend

The strategic reason they are doing this

The observability market is under pressure from three directions.

First, telemetry volumes are exploding. Kubernetes, microservices, edge, GPU clusters, AI workloads and distributed storage produce far more telemetry than humans can manually inspect.

Second, SRE teams are overloaded. Vendors are trying to sell “lower MTTR” and “less operational toil” by making the platform do more triage and correlation automatically.

Third, AI applications create new observability requirements. Traditional APM can tell you latency and error rate, but AI systems also need visibility into prompts, responses, hallucinations, drift, token usage, model quality, RAG retrieval quality, vector database behaviour and agent decisions.

So vendors are not just adding AI because it is fashionable. They are defending and expanding their core observability business.

What this means for an SRE / Observability Platform Engineer

The skill shift is significant.

Old value:

Build dashboards.
Write alert rules.
Know PromQL and LogQL.
Search logs manually.
Correlate incidents by experience.

New value:

Design telemetry that AI can reason over.
Standardise labels and service metadata.
Maintain accurate topology and ownership maps.
Connect observability to deployment and incident data.
Create safe remediation workflows.
Validate AI-generated RCA.
Control cost, access and blast radius.

The winners will not simply be the engineers who know the most dashboards. The winners will be the engineers who can build a trusted operational intelligence layer over metrics, logs, traces, topology and automation.

AI Strategies of New Observability Products

Coralogix is releasing the most explicit “AI observability product suite.” Cribl is positioning itself as the telemetry data layer for AI-era observability. Tsuga is newer and appears to be building an AI-native, bring-your-own-cloud observability architecture rather than simply adding an AI assistant to an old SaaS model.

Quick comparison

CompanyAI directionProduct maturity from public material
CoralogixAI Center, AI guardrails, AI evaluations, AI-SPM, Olly AI observability agentVery explicit productised AI offering
CriblCribl AI, Copilot, AI-guided Search Investigations, telemetry for humans and agentsStrong AI-assisted telemetry/data-management direction
TsugaBYOC observability for the AI era, agent-native observability, MCP/CLI for customer-owned agentsNewer; more architectural and agent-native positioning

1. Coralogix: AI observability as a full product suite

Coralogix is clearly releasing AI-focused products. Its main AI platform is AI Center, which Coralogix describes as a complete platform for AI-powered applications combining observability, guardrails, evaluations, and AI Security Posture Management in one place. It monitors LLM interactions for health, performance, cost, latency, errors, security and quality issues.

The key Coralogix AI products are:

Coralogix AI Center
├─ AI Observability
├─ AI Guardrails
├─ AI Evaluations
├─ AI Security Posture Management
├─ AI Application Discovery
└─ AI Explorer / Application Drilldown

What Coralogix is targeting

Coralogix is not just monitoring servers. It is monitoring AI application behaviour:

Prompt

LLM call

Response

Evaluation

Guardrail decision

Security / quality / cost signal

Its AI Center monitoring gives an organisation-level view of LLM usage and lets teams drill from a trend down to a specific application and even a specific prompt/response interaction.

It also supports OpenTelemetry GenAI semantic conventions, so teams can send GenAI spans into Coralogix AI Center without needing a Coralogix-specific SDK.

Olly: Coralogix’s AI observability agent

Coralogix also has Olly, which it describes as an AI-native observability agent. Olly lets users ask natural-language questions and get answers across logs, metrics, traces and alerts.

In practice, this is the “AI SRE assistant” layer:

Human asks:
“Why is payment latency rising?”

Olly checks:
├─ logs
├─ metrics
├─ traces
├─ alerts
├─ correlations
└─ possible root causes

Then returns:
├─ explanation
├─ evidence
├─ affected services
└─ recommended next steps

Coralogix also positions Olly as more than a simple assistant: it says Olly uses specialised agents for log analysis, trace exploration, metrics interpretation, security research, code debugging, correlation analysis and hypothesis generation.

My read on Coralogix

Coralogix is trying to own AI production reliability:

Monitor AI apps
Evaluate AI outputs
Detect prompt injection / PII / toxicity
Track token cost
Find bad model behaviour
Use AI to investigate normal production incidents

So yes: Coralogix is strongly AI-focused.

2. Cribl: AI platform for telemetry, not classic dashboard observability

Cribl’s AI angle is different. Cribl is not primarily trying to be another Datadog-style full-stack UI. It is positioning itself as the AI Platform for Telemetry: the collection, routing, shaping, searching and governance layer for machine data used by humans and AI agents. Cribl’s homepage describes the platform as giving enterprises choice and control for telemetry, and says it helps manage and analyse telemetry for both humans and agents.

The AI-focused Cribl areas are:

Cribl AI
├─ Copilot
├─ Copilot Editor
├─ AI-guided Search Investigations
├─ Natural-language queries
├─ AI-assisted pipeline creation
├─ AI telemetry parsing
└─ AI-ready telemetry routing

Cribl AI and Copilot

Cribl says its AI capabilities help teams create and modify pipelines, queries and configurations using natural language. It also says Cribl Copilot provides troubleshooting guidance, answers product/configuration questions and helps teams resolve issues faster.

This matters because a lot of observability toil is not just dashboards. It is:

Parse this log format.
Map this schema.
Route this data.
Drop this noisy field.
Mask this sensitive value.
Send this stream to the SIEM.
Send this other stream to cheaper storage.

Cribl’s AI is aimed at reducing that data-engineering toil.

Copilot Editor

Cribl’s Copilot Editor uses AI to help with schema mapping, translating logs across systems and building telemetry pipelines that clean, filter and route events.

That is important because AI-era observability needs clean, standardised telemetry. A reasoning agent is only useful if the data has usable structure.

Raw logs

AI-assisted parsing

Schema mapping

Enrichment / masking / routing

Search / SIEM / observability backend / AI agent

AI-guided Cribl Search Investigations

Cribl Search has an Investigations feature in preview. The docs describe it as a guided workspace where users explore incidents and telemetry using natural-language prompts. It helps analyse telemetry, identify patterns and document findings without manually building every query.

That means Cribl is moving into the AI-assisted investigation workflow:

Alert or question

Natural-language investigation

Generated queries

Pattern discovery

Findings captured in one workspace

Cribl’s AI observability thesis

Cribl’s recent AI observability messaging is that AI observability is a telemetry problem, not just a dashboard problem. It argues that LLM apps generate prompts, completions, tool calls, retrieval steps, token counts, model choices, policy events and infrastructure signals, and that those need to be collected and shaped for different teams and tools.

My read on Cribl

Cribl is not saying:

“We are the AI RCA dashboard.”

It is saying:

“We are the telemetry control plane that makes AI investigations possible.”

That is strategically clever. AI agents need cheap, governed, high-quality access to large telemetry volumes. Cribl wants to be the pipe, filter, schema and search layer underneath that.

3. Tsuga: AI-native observability architecture, still early

Tsuga is the newest and least mature publicly compared with Coralogix and Cribl, but it is very clearly positioning itself around the AI-era observability problem.

Tsuga describes itself as a bring-your-own-cloud observability platform for logs, metrics, traces and APM, deployed inside the customer’s AWS account using infrastructure-as-code. It says customers get the control of self-hosted infrastructure without the operational burden of running it.

Its newer positioning is explicitly AI-era focused. Tsuga announced a $35 million Series A on June 23, 2026, saying it is building “observability for the AI era” inside the customer’s cloud so the customer’s data and AI do not leave their control.

Tsuga’s AI claim

Tsuga’s argument is architectural:

Traditional observability:
telemetry leaves your cloud
vendor stores it
cost rises with volume
AI agents require broad access to vendor-hosted data

Tsuga model:
observability runs inside your cloud
telemetry stays inside your perimeter
AI runs on your own data
agents can use complete telemetry without exporting sensitive context

Tsuga says its AI tools run on the customer’s data inside the customer’s perimeter. It also says automated root-cause analysis runs on complete, unsampled data, and that its MCP server and CLI let engineering teams build their own agents on that foundation inside their own security boundary.

That MCP point is important. It suggests Tsuga is not only building an observability UI; it is exposing observability context to AI agents.

Agent-native observability

Tsuga has a specific Agent-Native Observability page. It says Tsuga is built so AI agents can use observability data effectively, affordably and inside the customer environment. It highlights agent-first APIs, MCPs, CLIs and query interfaces designed to return relevant context rather than raw data dumps.

That is a very modern product angle.

AI agent asks:
“What changed before this incident?”

Tsuga should return:
├─ relevant metrics
├─ relevant logs
├─ deployment context
├─ service ownership
├─ topology
└─ probable causal evidence

Not:
└─ 10GB of raw logs

What is less clear with Tsuga

Publicly, Tsuga looks less like:

Named AI assistant with lots of screenshots and feature modules

and more like:

AI-native observability architecture:
BYOC
complete telemetry
agent APIs
MCP
automated RCA
customer-owned AI boundary

So my assessment is: yes, Tsuga is AI-focused, but the public product story is currently more architectural and agent-native than feature-by-feature like Coralogix.

The strategic differences

Coralogix: “Observe and govern AI applications”

Coralogix is focused on production AI application reliability:

LLM monitoring
AI guardrails
Evaluations
AI security posture
Prompt/response visibility
Olly AI investigation agent

Best fit:

Teams deploying LLM apps and agents who need monitoring, safety, cost tracking and AI-assisted troubleshooting.

Cribl: “Prepare and control telemetry for AI”

Cribl is focused on the telemetry substrate:

Collect once
Shape data
Mask sensitive fields
Route anywhere
Search cheaply
Let humans and agents investigate
Use AI to build pipelines and queries

Best fit:

Large enterprises drowning in telemetry volume, SIEM costs, log routing complexity and multi-tool data sprawl.

Tsuga: “Run AI-era observability inside your own cloud”

Tsuga is focused on sovereign, cost-controlled, agent-native observability:

BYOC deployment
Telemetry stays in your cloud
AI and agents run inside your boundary
Automated RCA on unsampled data
MCP/CLI for custom SRE agents

Best fit:

Regulated, European, AI-native or high-scale companies that do not want telemetry, prompts, incident history and operational context exported to a third-party SaaS cloud.

The bigger market pattern

These newer players are attacking the incumbents from three angles:

1. Cost
AI generates more telemetry.
Per-GB SaaS observability becomes painful.

2. Data control
AI telemetry includes prompts, responses, business context and security-sensitive data.
Customers do not always want that in a vendor cloud.

3. Agent-readiness
Future observability is not just dashboards for humans.
AI agents need APIs, context retrieval, governed telemetry access and automated RCA.

So the new wave is less about “AI as a dashboard chatbot” and more about building the data foundation for AI-driven operations.

The sharpest summary is:

Coralogix = AI observability product suite
Cribl = AI-ready telemetry control plane
Tsuga = AI-native sovereign observability architecture

For an SRE/observability platform engineer, these companies are worth watching because they indicate where the next jobs and platform designs are going: telemetry engineering, AI-readable context, agent-safe access, automated RCA, guardrails and cost-controlled observability architectures.

Opensource AI Observability

AI adoption in open-source observability is happening, but it is different from what Datadog, Dynatrace, Splunk and New Relic are doing.

The commercial vendors are embedding AI directly into their SaaS platforms. The open-source ecosystem is mostly building the standards, collectors, SDKs, self-hostable platforms and agent interfaces that allow AI observability to work without vendor lock-in.

The big shift is this:

Old open-source observability:

Prometheus / Loki / Tempo / Grafana / OpenTelemetry

Collect, store, query, dashboard, alert


AI-era open-source observability:

OpenTelemetry + collectors + traces + logs + metrics + AI metadata

LLM / agent / RAG / GPU / vector DB visibility

AI assistants, AI SRE agents, natural-language querying, RCA

1. Grafana: open observability stack + AI features around it

Grafana Labs is moving in two directions.

First, it is keeping the open observability stack relevant for AI-era workloads: Grafana, Loki, Mimir, Tempo, Pyroscope and Alloy remain the core telemetry stack.

Second, it is adding AI-powered layers on top, especially in Grafana Cloud.

Grafana’s AI Observability product is built on OpenTelemetry and is aimed at teams running LLM agents in production. It monitors agent activity, traces conversations, tracks costs and evaluates quality. Grafana documents SDK support for Go, Python, TypeScript, Java and .NET, plus integrations with frameworks such as LangChain, LangGraph, OpenAI Agents and Vercel AI SDK.

Grafana also has Grafana Assistant, an AI-powered observability agent. It lets users ask questions like “Show me CPU usage” or “Create a dashboard for my database,” and it works across metrics, logs, traces, profiles and databases. Grafana says it can run investigations, manage dashboards, build/refine queries and help users navigate Grafana resources.

The important nuance: Grafana Assistant is not the same thing as open-source Grafana itself. It is primarily a Grafana Cloud AI capability, though Grafana documents a self-managed Assistant app that connects to a Grafana Cloud stack with reduced functionality.

Grafana’s most open-source-relevant AI move is probably Grafana Alloy. Alloy is Grafana Labs’ open-source OpenTelemetry Collector distribution with built-in Prometheus pipelines and support for metrics, logs, traces and profiles. It gives Grafana a standard collector layer for AI-era telemetry pipelines.

So Grafana’s strategy is:

Grafana AI strategy:

Open-source base:
Grafana
Loki
Mimir
Tempo
Pyroscope
Alloy

AI observability:
LLM / agent traces
cost tracking
quality evaluation
AI workload dashboards

AI assistant:
natural-language querying
dashboard creation
investigation assistance
query generation

Strategic direction:
keep the OSS stack open,
but place high-value AI workflows in Grafana Cloud.

2. OpenTelemetry: the standard layer for AI observability

OpenTelemetry is not a company; it is a CNCF open-source project. Its role is different from Grafana’s.

OpenTelemetry is becoming the standard telemetry schema and instrumentation layer for AI systems.

OpenTelemetry describes itself as an open-source observability framework for cloud-native software, providing APIs, libraries, agents and collector services for capturing telemetry. It also emphasises vendor-neutral instrumentation, meaning you instrument once and export to different backends.

For AI, the key development is OpenTelemetry semantic conventions for generative AI. OpenTelemetry has been extending its conventions so GenAI telemetry can capture model parameters, response metadata, token usage, traces, metrics and events for model interactions.

That matters because LLM systems need new telemetry fields that normal web apps did not need:

Traditional app telemetry:
service.name
http.status_code
duration
error
route
database call

AI app telemetry:
model name
prompt
completion
token count
tool call
retrieval step
vector DB query
embedding model
cost
temperature
hallucination score
safety evaluation

OpenTelemetry is not trying to become an AI assistant. Its value is that it gives the ecosystem a common language for AI telemetry.

So OpenTelemetry’s strategy is:

OpenTelemetry AI strategy:

Standardise:
spans
metrics
logs/events
attributes
semantic conventions

Support:
LLM calls
model interactions
prompts/responses
token usage
latency
errors
provider metadata

Enable:
Grafana
SigNoz
Langfuse
OpenLIT
Elastic
New Relic
Datadog
custom platforms

Strategic direction:
become the neutral telemetry contract for AI applications.

3. OpenLIT: open-source LLM observability on OpenTelemetry

OpenLIT is a good example of the new generation of open-source AI observability projects.

It describes itself as an open-source LLM observability and AI engineering platform built on OpenTelemetry. Its positioning is self-hosted, privacy-first and vendor-neutral.

This is important because many companies do not want prompts, responses, user inputs, sensitive data or AI-agent traces going straight into a third-party SaaS.

OpenLIT’s direction is:

OpenLIT strategy:

Monitor:
LLM calls
latency
token usage
cost
model behaviour
vector DBs
GPU usage

Deploy:
self-hosted
OpenTelemetry-native
privacy-first

Best fit:
teams building AI apps who want open-source AI observability
without committing to a commercial platform first.

4. Langfuse: open-source LLM tracing and evaluation

Langfuse is another major open-source AI observability project.

It focuses on LLM application tracing: capturing prompts, model responses, token usage, latency, tool calls and retrieval steps. Langfuse also provides AI-engineering features such as LLM-as-judge evaluation, prompt management, experiments and datasets, and it can be self-hosted.

Langfuse is less like “Grafana for all infrastructure” and more like “observability and evaluation for LLM applications.”

Its strategy is:

Langfuse strategy:

Trace:
prompt
response
tool call
RAG step
latency
token usage
cost

Evaluate:
quality
scoring
experiments
prompt versions
datasets

Best fit:
AI product teams who need to debug and improve LLM apps,
not just monitor infrastructure.

5. SigNoz: open-source observability with AI-agent access

SigNoz is moving from being an open-source Datadog/New Relic alternative into a more AI-aware observability platform.

SigNoz describes itself as an open-source observability tool powered by OpenTelemetry, covering logs, metrics, traces, dashboards, alerts and LLM/AI observability. It also advertises an MCP server for bringing telemetry into coding agents and an AI teammate called Noz for incident investigation, alert tuning and dashboard building.

This is significant because it shows a broader open-source pattern: observability platforms are not just adding AI dashboards; they are exposing telemetry to AI agents.

SigNoz direction:

OpenTelemetry-native observability
+
LLM/AI observability
+
MCP access for coding agents
+
AI teammate for investigations and dashboards

That is where open-source observability is going: not just dashboards for humans, but context APIs for agents.

6. HolmesGPT: open-source AI SRE agent

HolmesGPT is another important example because it is not primarily about observing LLM apps. It is about using AI to investigate production incidents.

HolmesGPT describes itself as an open-source AI agent for investigating production incidents and finding root causes across Kubernetes, VMs, cloud providers, databases and SaaS platforms. It is listed as a CNCF sandbox project.

That puts it closer to the Datadog Bits AI / Dynatrace Davis AI direction, but in open-source form.

HolmesGPT strategy:

Input:
alerts
Kubernetes state
metrics
logs
cloud context
runbooks

AI task:
investigate incident
gather evidence
find probable root cause
explain next action

Best fit:
platform teams wanting an open-source AI SRE layer
over existing observability tools.

The overall open-source adoption pattern

Open-source observability is adopting AI in four layers.

1. AI telemetry standards

This is where OpenTelemetry is most important.

Goal:
make AI applications observable in a standard way

Examples:
GenAI semantic conventions
token usage attributes
model request spans
prompt/response events
tool-call spans

This is foundational. Without standard AI telemetry, every vendor and OSS project invents incompatible schemas.

2. AI workload observability

This is where Grafana AI Observability, OpenLIT, Langfuse and SigNoz fit.

Goal:
monitor LLM apps, agents and RAG pipelines

Signals:
latency
token cost
prompt/response quality
hallucination risk
model errors
vector DB retrieval
tool calls
agent steps

3. AI-assisted operations

This is where Grafana Assistant, HolmesGPT, SigNoz Noz and similar tools fit.

Goal:
help humans investigate production systems faster

Capabilities:
natural-language querying
alert explanation
dashboard generation
root-cause hints
log summarisation
incident summaries

4. Agent-native observability

This is the newest layer.

Goal:
let AI agents consume observability data safely

Interfaces:
MCP servers
CLI tools
API access
context retrieval
guarded query execution
evidence-based RCA

This matters because future AI coding agents and SRE agents will need access to production telemetry to debug issues. The observability stack must become queryable by both humans and machines.

The key difference from commercial observability

Commercial vendors are building polished AI experiences inside their own SaaS platforms.

Open-source observability is building the portable foundations:

LayerOpen-source approach
InstrumentationOpenTelemetry SDKs and semantic conventions
CollectionOpenTelemetry Collector, Grafana Alloy
Storage/queryGrafana LGTM, SigNoz, ClickHouse-based stacks
AI app tracingOpenLIT, Langfuse, OTel GenAI conventions
AI SREHolmesGPT, MCP-enabled tools
Agent accessMCP, APIs, CLI workflows

The strategic difference is:

Commercial vendors:
"Use our platform and our AI will help you."

Open-source ecosystem:
"Instrument once, own your data, expose telemetry to any backend or AI agent."

What this means for SREs and observability engineers

The valuable skill is moving from only operating dashboards to building an AI-readable telemetry platform.

That means:

You need:
consistent OpenTelemetry attributes
clean service names
good resource metadata
deployment markers
trace/log/metric correlation
AI workload spans
token/cost metrics
evaluation signals
MCP or API access for agents
guardrails around sensitive telemetry

For a homelab or professional platform, the modern open-source direction would be:

Applications / AI agents

OpenTelemetry SDKs + GenAI semantic conventions

Grafana Alloy or OpenTelemetry Collector

Mimir / Loki / Tempo / ClickHouse / SigNoz / Langfuse / OpenLIT

Grafana dashboards + AI assistant / HolmesGPT / MCP-enabled agents

The sharp summary:

Grafana is making the open observability stack AI-aware.

OpenTelemetry is becoming the standard language for AI telemetry.

OpenLIT, Langfuse and SigNoz are making LLM apps observable.

HolmesGPT-style tools are turning open telemetry into AI-assisted SRE investigations.

So, yes: open-source observability is adopting AI quickly, but the centre of gravity is different. The open-source world is less about one vendor-owned AI brain and more about open telemetry, self-hostable AI observability, and agent-ready operations.