
From alert to root cause in 3 minutes: go.es.io/3UBe0Wv
✓Where developers learn, build, and share. Your source for hands-on demos, cheat sheets, explainers and more.

From alert to root cause in 3 minutes: go.es.io/3UBe0Wv

The Observability pillars (metrics, traces, and logs) are like tools to use, and you must use the right tool for the right job. The incident walkthrough starts with a checkout latency regression. p95 latency climbs from 180ms to over 2s and HTTP 500s appear. • Metric: the latency jump says something is wrong • Trace: a payment-gateway span grows from about 180ms to about 2500ms • Log: PoolExhaustedException, with the connection pool pinned at 20 of 20 The usual problem, per Elastic’s pillars post: “log data stores are siloed and not inherently correlated.” Elastic Observability’s answer is a single, integrated platform that combines logs, metrics, application traces and user experience data. In the Labs walkthrough, an agent correlates the three signals and opens a case.

Use your K8s PromQL unchanged in Elasticsearch Same query, no rewrite go.es.io/48oNSRN

The three pillars of Observabilty: go.es.io/4jg4AsM

Which of the three pillars of Observability offers you the right answer when an incident happens or a latency graph spikes? Three questions for an SRE or operations team: • Metrics: “what’s happening”. Numbers you can set alerting thresholds on • Traces: “where is it happening”. They help discover bottlenecks • Logs: “why is it happening”. Timestamped records of events Elastic Observability combines logs, metrics, application traces and user experience data in a single, integrated platform. The pillars post also proposed profiling as a fourth. OpenTelemetry Profiles reached Alpha in March 2026.


Elastic Workflows reduces the complexity of investigating and analyzing alerts. It ships an RCA example that does the boring half: • Attach it to an observability alerting rule • An Agent Builder agent investigates the alert and proposes root causes • A case opens with an AI-written title and description, the alert attached, the analysis as a comment • A final comment lists every reasoning step and tool call, so you can audit how it got there Deterministic steps do the plumbing. The agent only reasons where reasoning helps. In the Labs walkthrough the workflow ran in 3 minutes 22 seconds, all five steps successful. One run, manual trigger. The agent proposes root causes. The Labs walkthrough’s advice: start read-only. Self-managed, Elastic Workflows needs an Enterprise subscription. GA since 9.4. Docs linked below.

Reference docs: go.es.io/4frv4Fk

Deleting a document in Elasticsearch doesn't delete it. Not right away. Data lives in immutable Lucene segments. A delete or update only marks the old doc as gone. The disk space comes back when segments merge and get rewritten. Force merge does that rewrite on demand: > fewer segments, faster searches > deleted docs finally purged One catch: it's heavy disk and CPU work. Only run it on indices you've stopped writing to. In ILM, force_merge_index: true merges before the frozen snapshot, and that load lands on the tier before frozen. Tier struggling? Add capacity or a warm forcemerge phase. Don't just turn the merge off.

A cheap recall fix moved from step 31 to step 2. DiskBBQ auto-calibration walks a list of candidate configs in cost order and keeps the first one predicted to hit 90% recall@10. So the order of that list decides what you ship. The old order spent document bits too early. Rerank depth could swing the cost by more than a whole bit, so a 2-bit encoding with shallow reranking got tried before a 1-bit encoding with deeper reranking. Preconditioning only got a turn if every other config missed the target, which rarely happened. That matters because document bits are permanent: you pay for them on every vector for as long as the segment lives. Deeper reranking or preconditioning can buy the same recall without growing the index. We changed 2 things: • Cost model: 3.0 × document_bits + 1.3 × rerank_depth (the weight was 1), so every 1-bit config sorts before any 2-bit one • Preconditioning: tried on and off at every step, instead of as a fallback Across our 11 benchmark datasets, 4-bit symmetric encodings are gone and preconditioning now gets picked on 3 of them. The trade-off: where the config changed, recall@10 moved by at most 3 points in either direction with every cluster visited. Calibration still takes 1 to 2 seconds per merge. If auto-calibration is already on for your bbq_disk field, there's nothing to change. Newly merged segments pick up the new order, and existing ones keep their config until they merge. Full cost model, sweep order and per-dataset results: go.es.io/3WyRhuK

Elastic Streams is now a managed data pipeline for telemetry on Elastic Cloud Serverless. Your shippers keep sending as they do now. Streams filters, parses, enriches and routes the data to Elasticsearch, your own S3 bucket, or both. What that changes: - Full record in your S3 bucket, only the queried subset indexed - Every source, pipeline and destination in Streams Map - The whole config exports as YAML for pull requests - Elastic runs the OTel Collectors: nothing to size or patch Private preview opens today for a selected group of design partners: go.es.io/3Vhu0wY

You rotated your CA. Every config file has the new cert. Some clients still fail with: x509: certificate signed by unknown authority The culprit: long-running processes. Most clients load their trust store once, at startup. Changing the file on disk changes nothing until they restart. We've seen Metricbeat sidecars up for months still trusting the old CA. Safer rotation: > Trust both old + new CA first > Restart every client, sidecars included > Then switch the server cert > Remove the old CA last


Reference docs: go.es.io/3UkQVHp

Your API's average latency is 80ms. Looks fine. Now look closer: 98 requests took 20ms. 2 took 3 seconds. Averages hide the tail. And the tail is real users, often your heaviest ones, because more requests means more chances to hit it. Watch p95 and p99, not just the mean. In Elastic APM, switch the latency chart from Average to 95th or 99th percentile and see who's actually waiting.

Reference docs: go.es.io/4iLGHJf