Monitoring thresholds

Reference thresholds for the per-service alerting required by Production Readiness (M1 to M7). Adapt the exact values to your capacity and criticality, but keep the two levels: warning (working hours) and critical (24/7 when the application is critical).

PostgreSQL / MongoDB

Metric	Warning	Critical
Availability (health check)	-	failure > 1 min
No primary (Mongo) / standby unavailable (PG)	-	> 30 s
Replication lag	> 30 s	> 5 min
Connections used / max_connections	> 70%	> 85%
`remaining connection slots are reserved` error	-	any occurrence
Sustained CPU	> 70% for 15 min	> 90% for 5 min
Available memory	< 25%	< 10%
Disk used	> 75%	> 85% or predicted saturation < 7 days
IOPS / burst credits	> 80% consumed	imminent exhaustion
Query latency p95	drift > 2x baseline	-
Long queries / transactions, deadlocks	> 5 min / deadlock spikes	-
Cache hit ratio (PG)	< 95%	-
Autovacuum behind / XID age (PG)	> 1.0 billion	> 1.5 billion
WiredTiger tickets available (Mongo)	< 30	< 10
Last successful backup	> 24 h	> 48 h or job failure
TLS certificate	expiry < 30 days	expiry < 7 days
PgBouncer clients waiting (cl_waiting)	> 0 for 1 min	pool saturated persistently

Redis

Metric	Warning	Critical
Availability / PING	-	failure > 30 s
used_memory / maxmemory	> 75%	> 90%
`OOM command not allowed` errors (noeviction instances)	-	any occurrence
evicted_keys on a noeviction instance	-	> 0 (misconfiguration)
Connected clients / maxclients	> 70%	> 85%
blocked_clients, rejected_connections	> 0	sustained growth
Command latency p99 / slowlog	> 10 ms	> 100 ms
Replication link broken / failover	-	any occurrence
rdb_last_bgsave_status / aof_last_write_status	-	err
Broker stream length (XLEN) vs maxLen	> 70%	> 90%
Consumer group lag (XPENDING)	continuous growth > 15 min	business threshold exceeded
Memory fragmentation (mem_fragmentation_ratio)	> 1.5	> 2.0

The broker’s consumer group lag is the most useful leading indicator: events accumulate before users perceive anything.

Elasticsearch / OpenSearch

Metric	Warning	Critical
cluster.health.status	yellow > 15 min	red (any occurrence)
unassigned_shards	> 0	> 0 for 15 min
Shards per node	> 800 (80% of the limit)	> 950
JVM heap used	> 75%	> 85%
Old-gen GC pauses	> 1 s cumulative per min	increasing frequency
Disk used	> 80% (before low watermark 85%)	> 88%; at 95% indices go read-only
Write / search threadpool rejections	> 0	sustained growth
Indexing / search latency p95	drift > 2x baseline	-
Circuit breakers triggered	> 0	sustained
Last snapshot failure / age	> 24 h	> 48 h or failure
ILM errors / policies in error	> 0	-
`cleanup-*` CronJob failures	1 failure	2 consecutive failures
Mapping errors (`Limit of total fields`, `mapper_parsing_exception`) in `prismeai-events` logs	any occurrence	recurrence
Node loss / master quorum loss	-	any occurrence
Disk iowait	> 10%	> 25%

Object storage

Metric	Warning	Critical
5xx error rate	> 0.1%	> 1%
403 / AccessDenied error rate	abnormal spike (expired credentials)	-
Request latency p95	drift > 2x baseline	-
Cross-region replication lag / failure	> 1 h	> 6 h
Change to bucket policy, public access block, or versioning	-	any occurrence (security alert)
Volume growth	drift > 30% vs monthly forecast	-
Age of last backup sync	> 24 h	> 48 h
Provider request quotas / limits	> 80%	> 90%

RWX filesystem

Metric	Warning	Critical
Space used	> 75%	> 85%
Inodes, burst credits / throughput (EFS), NFS latency, pod mount failures	to monitor	per runbook

Overview

Installation

Databases

Production Readiness

Entreprise Services

Operations

Monitoring thresholds

PostgreSQL / MongoDB

Redis

Elasticsearch / OpenSearch

Object storage

RWX filesystem

References

Production Readiness

Scaling

​PostgreSQL / MongoDB

​Redis

​Elasticsearch / OpenSearch

​Object storage

​RWX filesystem

​References

Production Readiness

Scaling

PostgreSQL / MongoDB

Redis

Elasticsearch / OpenSearch

Object storage

RWX filesystem

References