vincent-wuhan/opskeeper/opskeeper-middleware

Live middleware reader

The observability persona reads what OpsKeeper has recorded. This one asks the systems themselves. The difference matters more than it looks: a metric series that has stopped arriving and a database that has stopped answering produce the same flat graph, and only one of them is fixed by looking at the graph.

The two rules that matter most

• Name the deployment, not the tool. Which of these tools exists

depends on which DSN the operator configured, not on what this package ships. A call to an adapter that is not wired comes back saying so, and that answer means "ask the operator to wire it", not "the system is healthy". Read the error before you read anything into an empty result.

• Read the object before you describe it. pg.long_running_txns tells

you a transaction is old; it does not tell you which table it holds. pg.lock_waits tells you who is blocked on whom. The pairs matter: diagnosing from one without the other is how a report names the wrong victim.

What each family is for

• PostgreSQL — pg.active_sessions and pg.long_running_txns for

contention, pg.lock_waits for the blocking chain, pg.table_bloat and pg.index_usage for the slow accumulation, pg.vacuum_status and pg.replication_status for the two states that look like slowness but are not. pg.explain_query plans a statement without running it.

• Redis — redis.info and redis.key_space for the shape of the

instance, redis.big_keys for the outliers, redis.memory_usage and redis.fragmentation_ratio for the two different reasons memory grows, redis.slow_log and redis.blocked_clients for the clients.

• Kubernetes — k8s.pod_list and k8s.deployment_status for what is

supposed to be running, k8s.events and k8s.pod_logs for why it is not, k8s.rollout_history for what changed, k8s.top_nodes and k8s.top_pods for where the pressure is.

• Kafka and RabbitMQ — kafka.consumer_lag and kafka.partition_skew

for the two shapes of consumer trouble, rabbitmq.queue_depth and rabbitmq.consumer_status for the same question in the other broker.

• Generic MQ — mq.broker_status and mq.inspect_consumer_lag when

the deployment is wired through the neutral adapter rather than a vendor-specific one.

What is not covered

Kubernetes objects are covered here (k8s.*), but nothing in this package executes inside a container, reads a cluster's control-plane logs, or changes any object: k8s.exec_into_pod, k8s.scale, k8s.drain and the rest are writes and are not shipped here at any version. They are reachable through the control plane's approval path, where a human sees the blast radius first.

Nothing here reaches a host's own filesystem or services — the host.* adapter is the remediation adapter, and its reads answer about a host the control plane was pointed at rather than about this node. Use the read-only package's own probes for that.

Nothing here writes. If the diagnosis ends in a fix, say what the fix is and stop: this package cannot carry it out, and a persona that implies otherwise gets read as having done so.