34.1 第一动作：流量型还是保留型

SELECT application_name,
       state,
       wait_event_type,
       wait_event,
       count(*) AS sessions,
       max(clock_timestamp() - coalesce(xact_start, query_start))
         AS oldest_age
FROM pg_stat_activity
WHERE backend_type = 'client backend'
GROUP BY 1, 2, 3, 4
ORDER BY sessions DESC;

application arrival / timeout / retry
pool waiting clients / server connections
proxy accept / queue / backend health
PostgreSQL active / idle-in-transaction / Lock waits
host run queue / memory pressure / device latency

arrival/retry increases
  -> admission or execution concurrency increases
  -> queue/wait/rejection grows
  -> completion rate or user success falls

SELECT pid,
       usename,
       application_name,
       state,
       xact_start,
       backend_xid,
       backend_xmin
FROM pg_stat_activity
WHERE backend_xid IS NOT NULL
   OR backend_xmin IS NOT NULL
ORDER BY xact_start NULLS LAST;

SELECT slot_name,
       slot_type,
       active,
       xmin,
       catalog_xmin,
       restart_lsn,
       wal_status,
       safe_wal_size,
       invalidation_reason
FROM pg_replication_slots
ORDER BY slot_name;

SELECT transaction, gid, prepared, owner, database
FROM pg_prepared_xacts
ORDER BY prepared;

t0: free bytes, current LSN, restart_lsn, archive failure count
t1: same fields after a known interval

retained bytes     = current LSN - restart_lsn
growth rate        = (retained_t1 - retained_t0) / elapsed
time to hard limit = usable headroom / positive growth rate

flow:
  stop admission -> cancel exact work -> queue falls

retention:
  preserve owner/lineage evidence -> repair consumer/archive
  -> only then release an exact, authorized boundary

classification:
  flow: confirmed
  retention: confirmed
immediate_hard_failure: filesystem-full-in-18m
parallel_controls:
  - reduce noncritical write admission
  - preserve slot/archive evidence and contact owner
forbidden:
  - delete pg_wal files
  - drop unknown slot

which path is failing?
which object is growing?
who owns the oldest xmin/restart_lsn?
is there a valid backup or replica copy?
what client work will be canceled?
what data or recovery capability can be lost?

UTC + monotonic timestamp
user-facing probe and exact error class
pg_stat_activity grouped by app/state/wait
blocker tree and long/prepared transactions
pg_replication_slots and pg_stat_archiver
current/replay LSN and write rate
filesystem by mount/path plus inode usage
CPU/memory/pressure/device queue
recent config/deploy/failover/backup changes

FLOW
  sufficient connection/queue/wait evidence
  and no contradictory retention evidence for this action

RETENTION
  exact owner boundary and retained quantity observed
  and flow relief cannot release that boundary

STOP_AND_INVESTIGATE
  both or neither predicate is sufficiently proved

symptom: exact user and resource observation
scope: service / database / node / time window
classification: FLOW | RETENTION | BOTH | UNKNOWN
supporting_evidence: [...]
contradicting_evidence: [...]
first_action: bounded and reversible
stop_condition: measurable
owner: named role