Skip to content

How AlertBird works

AlertBird runs in your Kubernetes cluster. It turns received OpenTelemetry data into a service map, service and connection measurements, grouped errors and direct alerts. PostgreSQL stores compact history, bounded evidence, configuration and durable delivery work. The web app provides Service health, Alerts, Data & retention, and Integrations pages.

Collection and analysis

Upstream OpenTelemetry handles collection, resource enrichment, bounded queues, transport and initial CLIENT/SERVER span pairing. AlertBird uses the stock Collector; there is no proprietary Collector distribution or AlertBird Operator.

AlertBird handles canonical request, error, HTTP-status and latency aggregation from unique supported SERVER spans, service/connection health, coverage, error grouping, prospective log rules, retention, alert evaluation and delivery. Analysis uses supported received data before diagnostic evidence reduction. It cannot recover requests filtered, sampled or lost before they reach it.

Signal What it establishes
SERVER spans Canonical per-service request/error/status/latency measurements. HTTP 5xx defines canonical HTTP errors.
Upstream servicegraph output Connection observations and paired client/server latency. Connector failures use OTel ERROR status.
Supported native metrics Separate infrastructure/runtime measurements, with source and unit semantics preserved.
Exception spans/logs Error observations and bounded grouped evidence.
Enabled log matchers Counts of matching arrivals after the rule was enabled; retries may repeat arrivals.
Selected diagnostic traces Evidence for investigation, never extra requests.

Graph observations, native metrics and retained errors never add to the canonical request denominator. Summing every service's SERVER spans also counts internal hops; it is not a distinct ingress-request count.

Two installation modes

Included collection installs node collectors for pod logs and kubelet metrics in one selected workload namespace, a singleton gateway and install-or-reuse upstream Operator support. The Operator manages Collector custom resources. Application auto-instrumentation is separately opted in for supported runtimes; installing the chart does not instrument every workload.

Existing telemetry installs the same gateway as a plain Deployment. Existing Collectors or SDKs export to it. This mode installs no node agents, Operator, OTel CRDs or injection resources.

Both modes include the API, worker, app and bundled persistent or external TLS PostgreSQL. Helm manages release configuration. There is no cluster-administration UI. See quickstart and collection.

Health includes uncertainty

Service discovery follows observed instrumented traffic. Missing traffic alone does not establish availability. Current receipt freshness and historical window closure are separate checks. Numeric rules need supported source-quality assumptions, observations in the evaluated window, a closed window, fresh receipt times and no invalidating loss/sampling/reset reasons.

Unknown, stale and partial coverage stays visible. Declaring an unsampled source does not prove instrumentation or turn old gaps into healthy data. The service map is a view of observed scope, not a complete inventory of a cluster.

Operational bounds

The initial installation supports one organization/cluster and one pairing gateway. The gateway uses one replica and Recreate rollout; overlapping pairing instances and horizontal gateway scaling are unsupported. The packaged graph store holds 1,000 pending pairs with a 10-second TTL. Export queues and pairing state are bounded and held in memory. Restart can lose buffered data, pending pairs and unread logs; transport acknowledgement is not an end-to-end durability guarantee.

The backend uses explicit admission, cardinality and evidence bounds. Chart resource limits are a starting envelope, not a throughput or seven-day capacity promise. Monitor actual pipeline coverage, database growth and worker progress. See retention for the storage boundary.

Alerts go directly to configured providers from the worker, with durable state in PostgreSQL and no required AlertBird cloud relay. Provider acceptance and human receipt are separate outcomes. Run independent failure monitoring outside the installation. See alerting.

A full trace/log lake, general query builder, dashboard builder, on-call schedule and native desktop client are outside the current product scope. Repository access, licensing and production operation remain separate from local technical validation.

View Markdown source