marten-pg-resilience

A Crystal shard that patches crystal-pg to retry idle-killed TLS connections instead of surfacing a 500 error.

The problem

crystal-pg (as of 0.29.0) only rescues IO::Error when re-raising as DB::ConnectionLost so the connection pool can retry on a dead socket. When PostgreSQL is reached over a TLS connection (managed Postgres, Tailscale, any firewall that reaps idle connections), a half-closed socket surfaces as OpenSSL::SSL::Error ("SSL_write: Unexpected EOF while reading"), which is not an IO::Error — it inherits from OpenSSL::Error -> Exception. The exception bubbles all the way out and the first request after an idle period 500s instead of being transparently retried on a fresh connection.

The fix

This shard reopens PG::Statement#perform_query and #perform_exec to also rescue OpenSSL::SSL::Error and raise DB::ConnectionLost, enabling the pool to retry. Production-proven on a Writebook Marten deploy.

Remove this shard once crystal-pg upstreams the fix.

Installation

Add to your shard.yml:

dependencies:
  marten_pg_resilience:
    github: stevegeek/marten-pg-resilience

Then require it after your pg require (typically in your app entrypoint):

require "marten_pg_resilience"

No other configuration is needed. The patch is a no-op when not using the PostgreSQL backend (e.g., SQLite in development/test).

Provenance

Extracted from a production Writebook Marten port.

License

MIT