I am trying to deploy an automated high-available PostgreSQL cluster on kubernetes. In cases of master failover or temporary failures in master, standby loses streaming replication connection and when retrying, it takes a long time until it gets failed and retries.
I use PostgreSQL 10 and streaming replication (cluster-main-cluster-master-service
is a service that always routes to master and all the replicas connect to this service for replication). I've tried setting configs like connect_timeout
and keepalive
in primary_conninfo
of recovery.conf
and wal_receiver_timeout
in postgresql.conf
of standby but I could not make any progress with them.
In the first place when master goes down, replication stops with the following error (state 1):
2019-10-06 14:14:54.042 +0330 [3039] LOG: replication terminated by primary server
2019-10-06 14:14:54.042 +0330 [3039] DETAIL: End of WAL reached on timeline 17 at 0/33000098.
2019-10-06 14:14:54.042 +0330 [3039] FATAL: could not send end-of-streaming message to primary: no COPY in progress
2019-10-06 14:14:55.534 +0330 [12] LOG: record with incorrect prev-link 0/2D000028 at 0/33000098
After investigating Postgres activities I found out that WalReceiver proccess stucks in LibPQWalReceiverConnect
wait_event (state 2) but timeout is way longer than what I configured (although I set connect_timeout
to 10 seconds, it takes about 2 minutes). Then, It fails with the following error (state 3):
2019-10-06 14:17:06.035 +0330 [3264] FATAL: could not connect to the primary server: could not connect to server: Connection timed out
Is the server running on host "cluster-main-cluster-master-service" (192.168.0.166) and accepting
TCP/IP connections on port 5432?
In the next try, It successfully connects the primary (state 4):
2019-10-06 14:17:07.892 +0330 [5786] LOG: started streaming WAL from primary at 0/33000000 on timeline 17
I also tried killing the process when stuck event occurs (state 2), and when I do, It starts the process again and connects and then streams normally (jumps to state 4).
After checking netstat, I also found that there is a connection with SYN_SENT
state to the old master in the walreceiver process (in failover case).
connect_timeout
governs how long PostgreSQL will wait for the replication connection to succeed, but that does not include establishing the TCP connection.
To reduce the time that the kernel waits for a successful answer to a TCP SYN request, reduce the number of retries. In /etc/sysctl.conf
, set:
net.ipv4.tcp_syn_retries = 3
and run sysctl -p
.
That should reduce the time significantly.
Reducing the value too much might make your system less stable.