-
Type:
Problem report
-
Resolution: Unresolved
-
Priority:
Major
-
None
-
Affects Version/s: 7.0.32, 8.0.0
-
Component/s: Proxy (P), Server (S)
-
Environment:DB connections through HAProxy or other load balancers (Azure, Amazon, AliCloud), PG-pool, Pg-Bouncer, MaxScale, etc...
-
Support backlog
Steps to reproduce:
Server and proxy have no read/write timeout on SQL backend connections. If a proxy between Zabbix
and the database (HAProxy, ProxySQL, PgBouncer) loses its backend but keeps the client TCP session
open, the Zabbix process waits forever in recv(). TCP keepalive doesn't help, because the proxy's
kernel answers the probes. Same class of problem as ZBX-28358 (ClickHouse/Elasticsearch).
Example:
- Galera/Patroni cluster behind HAProxy (single-writer); Zabbix server connects via HAProxy.
- Force writer switchover while history syncers are mid-query (e.g. freeze a node briefly).
- HAProxy drops the backend session but keeps the frontend session open.
Result:
History syncers hang indefinitely.
No "database is down" message, no reconnect.
Expected:
The query fails after a configurable timeout, the connection is treated as down, and the server reconnects.
- part of
-
ZBX-28358 Enable clickhouse and elasticsearch keepalive and timeouts
-
- READY TO DEVELOP
-