pg0 0.14.1 start silently hangs (0% CPU, no output) on existing instance on Windows

[bchumerch]

Summary

On Windows, pg0 start --name <existing-instance> silently hangs: the pg0 process is alive at ~0% CPU, never writes anything to stdout or stderr, never binds the port, never creates postmaster.pid. Fresh instances (data dir doesn't yet exist) work perfectly, so the binary itself is fine — only the existing-instance path hangs.

Workaround is calling pg_ctl start directly with the exact arguments pg0 -v start uses internally (captured from a verbose run on a fresh instance). I'm filing this so the underlying behavior can be fixed.

Environment

• pg0: 0.14.1 (pg0-embedded pip package), binary dated 2026-05-08
• OS: Windows 11 Home 10.0.26200
• PostgreSQL: 18.1.0 (bundled)
• Installation root: C:\Users\<user>\.pg0\installation\18.1.0
• Instance affected: any pre-existing instance with a populated data dir

Reproduction

1. Create a fresh instance and let pg0 initialize it:

   pg0 -v start --name pg0-test --port 5444
   

   This works — full verbose output, postgres binds in ~7s, instance.json and postmaster.pid created.

2. Stop the fresh instance, leaving the data directory intact:

   pg0 stop --name pg0-test
   

3. Restart the existing instance:

   pg0 -v start --name pg0-test
   

   pg0 process spawns, sits at 0% CPU for minutes, produces zero bytes to stdout/stderr, never binds the port. Eventually it can be killed; no error code is observable because nothing was emitted.

Evidence

• Get-Process pg0 shows the process alive with CPU=0.015 (i.e. essentially nothing) after 3+ minutes.
• Get-NetTCPConnection -LocalPort <port> returns nothing.
• The redirected log files captured by the parent process (pg0.out.log, pg0.err.log) are 0 bytes at rest.
• pg0 info --name <instance> returns "stopped" status correctly — so pg0 can read instance.json; it just hangs on the start path.

For comparison, the fresh-instance verbose run logs everything:

Setting up PostgreSQL 18.1.0...
DEBUG pg0: PostgreSQL already extracted at C:\Users\<user>\.pg0\installation\18.1.0
DEBUG setup: initialize: postgresql_embedded::postgresql: Initializing database <path>
DEBUG setup: initialize: execute_command{program="initdb"}: ...
[... initdb output ...]
DEBUG start: postgresql_embedded::postgresql: Starting database <path> on port <port>
DEBUG start: execute_command{program="pg_ctl"}: ...
DEBUG start: postgresql_embedded::postgresql: Started database <path> on port <port>

For the hanging existing-instance case, none of these lines appear — not even the first Setting up PostgreSQL 18.1.0... line. So the hang is before the "setup" tracing span enters.

What was tried

• Pid value in instance.json does not matter — tested with pid: 0 (Windows System Idle Process, always alive) and pid: 999999 (definitely dead). Both hang identically.
• Stale postmaster.pid removal doesn't help — hang occurs whether or not postmaster.pid is present.
• Same hang reproduces against two different existing instances (different data dirs, different ports).
• Killing any zombie pg0 processes from prior attempts first, also no help.

Workaround that works

Run postgres directly via the bundled pg_ctl, using the exact args pg0 would have used (captured from a fresh-instance verbose run):

pg_ctl start --pgdata <data> --wait --log <data>/start.log \
  -o "-F -p <port>" \
  -o "-c log_directory=log" \
  -o "-c timezone=UTC" \
  -o "-c work_mem=64MB" \
  -o "-c log_rotation_size=100MB" \
  -o "-c maintenance_work_mem=512MB" \
  -o "-c shared_buffers=256MB" \
  -o "-c log_timezone=UTC" \
  -o "-c log_rotation_age=1d" \
  -o "-c effective_cache_size=1GB" \
  -o "-c max_parallel_maintenance_workers=4" \
  -o "-c logging_collector=on" \
  -o "-c log_filename=postgresql-%Y-%m-%d.log"

Crash recovery completes in <100ms, postgres binds the port, everything works.

Suggested investigation

The hang is happening before any of pg0's tracing spans enter, suggesting it's during instance-state reading / lock acquisition / Windows file-system call rather than in the postgres invocation itself. Possible suspects:

1. A blocking Windows process-introspection call (e.g. checking if an old pid is alive)
2. A file lock or named-pipe wait inherited from a prior crash
3. A network/DNS lookup somewhere in the supervisor

The fact that fresh instances work but any restart against a populated data dir hangs suggests the diverging code path is the "I already have an instance.json and a data dir — let me figure out what state postgres is in" branch.

Related closed issues

• #6 — stale postmaster.pid path (data deletion). Different symptom (data loss vs silent hang), but adjacent code area. The fix for pg0 start deletes data directory on stale postmaster.pid (data loss after unclean shutdown) #6 may have introduced this new branch.
• #13 — localized PG error matching on restart. Same surface (start an existing instance on Windows), different failure mode.

Happy to capture additional traces or test a patch.

[nicoloboschi]

Resolved by #27: Windows now uses direct process-handle checks instead of the blocking tasklist probe when starting an existing instance.