[-] fix PgBouncer reconnect by pgxpool ping, closes #1570

[pashagolub]

Description

pgwatch reconnects to PgBouncer on every metric fetch. PgBouncer logs an error each time:

ERROR invalid command '-- ping', use SHOW HELP;
WARNING C-0x...: pgbouncer/pgbouncer@...:36340 pooler error: invalid command '-- ping', use SHOW HELP;
LOG C-0x...: pgbouncer/pgbouncer@...:36340 closing because: client close request (age=5s)
LOG C-0x...: pgbouncer/pgbouncer@...:36354 login attempt: db=pgbouncer user=pgbouncer

Cause

pgxpool.Acquire pings any connection that has been idle for more than a second before handing it out (pgxpool/pool.go), and pgconn.Ping sends -- ping. The PgBouncer admin console rejects that statement, so the pool discards the connection and opens a new one. With pgbouncer_stats: 5 this happens on every fetch.

This is not a regression. pgxpool has pinged on acquire since pgx v5, and source connections have been pools since #521. #639 fixed the same error message for the connect path only, by replacing pgwatch's own Conn.Ping call with SHOW VERSION; the pool-internal ping was never covered, so #625 was closed while half-fixed.

Fix

Set ShouldPing to always return false for PgBouncer sources. Connect() and Ping() still validate the connection with SHOW VERSION, so liveness is checked where it matters; a connection that dies while idle now surfaces as an error on the next metric query instead of at acquire time. Telegraf took the same approach in influxdata/telegraf#19258.

ShouldPing requires pgx v5.7.6 or newer; we are on v5.11.0.

Verification

Reproduced the bug against cybertecpostgresql/pgwatch:6.0.0 and edoburu/pgbouncer:v1.25.2-p0 with the compose file from the issue: a fresh login every 5 seconds, each closing with invalid command '-- ping'.

The fix itself has not yet been run against a live PgBouncer; it is verified by code inspection only. Worth a manual check before merge.

Fixes #1570

AI & Automation Policy

• I am the human author and take full personal responsibility for every change in this PR.
• No AI or automated generative tool was used in any part of this PR OR I have disclosed all tool(s) below.

AI/automation tools used (leave blank if none):

Investigated, reproduced and drafted with Claude Code.

Checklist

• Code compiles and existing tests pass locally.
• New or updated tests are included where applicable.
• Documentation is updated where applicable.

go build ./internal/sources/ passes and gofmt is clean, but the full build and test run could not be done locally: protoc is not installed, so api/pb is not generated and internal/testutil fails to compile. Leaving the first box for CI and the author.

[coveralls]

Coverage Report for CI Build 35357768818

Coverage decreased (-0.1%) to 87.777%

Details

• Coverage decreased (-0.1%) from the base build.
• Patch coverage: 2 of 2 lines across 1 file are fully covered (100%).
• 8 coverage regressions across 1 file.

Uncovered Changes

No uncovered changes found.

Coverage Regressions

8 previously-covered lines in 1 file lost coverage.

File	Lines Losing Coverage	Coverage
internal/sinks/prometheus.go	8	92.19%

---

Coverage Stats

Relevant Lines:	6193
Covered Lines:	5436
Line Coverage:	87.78%
Coverage Strength:	1.0 hits per line

---

💛 - Coveralls