Monitoring a Data Platform: What Should You Actually Measure?

Monitoring a Data Platform: What Should You Actually Measure? is a production engineering concern, not just an implementation detail. The useful design connects user intent to controlled execution, measurable outcomes, and evidence that remains available during an incident.

Monitoring layers

Follow the request from User or Client to Application, API, Report Engine, Data Platform, and finally Compute and Storage. Keep one correlation identifier across those boundaries. The client view supplies perceived latency; the application supplies feature and error context; the API supplies rate and status; the report engine separates validation, cache, and query work; the platform exposes SQL, pipeline, and notebook execution; compute and storage expose saturation and I/O.

Evidence types

Metrics are numeric time series suited to aggregation and alerts. Logs describe individual occurrences. Events record meaningful changes such as deployments or retries. Traces connect timed spans across services. Audit data records who requested or changed what. They have different retention and access needs, but shared identifiers and UTC timestamps make them useful together.

What to measure

Measure request volume, error rate, P50, P95 and P99 latency, concurrent requests, queue depth, cache hit rate, CPU, memory, reads, writes, waits, query duration, pipeline duration, notebook duration, failures, retries, rows processed, and data freshness. Pair every rate with a denominator and every duration with workload size and class.

Why averages fail

If most requests finish in milliseconds while a few take tens of seconds, the average describes almost nobody. P50 shows typical behavior, P95 exposes the slow edge, and P99 reveals rarer tail behavior. Include sample count and segment by report type, cache outcome, tenant class, result size, and success state.

Production considerations

Treat tenant isolation as a first-class boundary. Derive tenant identity from authenticated claims or a trusted service mapping, never only from tenant_id supplied by a browser. Apply tenant filters in the controlled query layer, include tenant identity in cache keys, protect stored results, and write an audit event for access. Redact secrets and sensitive filter values from ordinary logs.

Every operation needs a request or execution ID plus a correlation ID that crosses service boundaries. Capture UTC timestamps, status, duration, workload size, and error code. Keep high-cardinality detail in logs or traces rather than unbounded metric labels. Make telemetry asynchronous and bounded so a monitoring outage cannot take down the production path.

Define limits before scale exposes missing policy: maximum date range, maximum rows and bytes, execution timeout, concurrency per tenant, queue capacity, retry budget, and artifact retention. Reject invalid work early with a specific response. Retry only transient operations and use idempotency keys where duplicate execution could create extra files or charges.

Validate with representative data and failure drills. Test empty results, boundary dates, invalid dimensions, cross-tenant attempts, dependency timeouts, cache corruption, cancellation, retries, and large outputs. Compare the visible result with source totals and retain enough context to reproduce the decision. Operational readiness means an on-call engineer can identify the failing layer and take a bounded action without guessing.

Tradeoffs and failure modes

More telemetry improves diagnosis but adds storage, privacy, and cardinality costs. More caching reduces query load but creates freshness and invalidation risks. More flexible requests improve usefulness but expand the security and performance surface. Prefer explicit report or execution contracts, allow-listed variation, and measured exceptions over an unrestricted interface.

Watch for partial success: a pipeline can write data and fail before logging completion; a report can finish after its caller disconnects; a cache write can fail after a valid result was returned. Model those outcomes explicitly. Do not relabel an unknown value as zero, and do not overwrite failed attempts when a retry succeeds. Preserve the original error even if secondary logging also fails.

Practical rollout

Begin with one important workload and a small set of service objectives. Instrument the complete path, establish a baseline, and review evidence with application, data, security, and operations owners. Add alerts only when the receiver has a documented response. Expand by workload class after identifiers, access controls, and retention have proved reliable. This creates an operating model, not merely a dashboard.

Operational review questions

Before release, ask whether a responder can identify the affected client, tenant, workload class, code version, and dependency from retained evidence. Confirm that success means the intended data was delivered, not merely that a process exited without an exception. Check whether a retry is safe, whether cancellation stops downstream work, and whether partial output can be mistaken for a complete result.

During review, compare normal, peak, and failure behavior with representative volume. Verify that limits produce explicit outcomes and that dashboards distinguish rejected, queued, executing, completed, failed, and cancelled work. Assign ownership for the service, data contract, alerts, cache or operational store, and recovery procedure. Record decisions close to the implementation so future changes preserve the reasoning. Finally, test the investigation path with someone who did not build the feature; if that person cannot move from symptom to a specific execution and dependency, the design still lacks operational context.