Hasura 2.42.0 regresssion connectivity to PostgreSQL read only db

[SzymonZy]

Version Information

Server Version: 2.42.0
CLI Version (for CLI related issue):

Environment

OSS

What is the current behaviour?

I'm attaching hot standby postgresdb and hasura fails with:
{"detail":{"http_info":{"content_encoding":null,"http_version":"HTTP/1.0","ip":"172.28.63.18","method":"POST","status":500,"url":"/v2/query"},"operation":{"error":{"code":"postgres-error","error":"postgres tx error","internal":{"arguments":[],"error":{"description":null,"exec_status":"FatalError","hint":null,"message":"cannot set transaction read-write mode during recovery","status_code":"0A000"},"prepared":true,"statement":"BEGIN ISOLATION LEVEL READ COMMITTED READ WRITE"},"path":"$"},"query":{"args":{"source":"Replicated","sql":"SELECT VERSION()"},"type":"run_sql"},"query_execution_time":6.270142e-3,"request_id":"6a78c842-bd77-4c22-b33a-1da06841ebcb","request_mode":"error","request_read_time":1.41e-6,"response_size":327,"uncompressed_response_size":327,"user_vars":{"x-hasura-role":"admin"}},"request_id":"6a78c842-bd77-4c22-b33a-1da06841ebcb"},"level":"error","span_id":"d87a0e9b79825cc4","timestamp":"2024-11-05T10:45:43.256+0000","trace_id":"3f335bf56449eb9f5cc0fe745c98ba19","type":"http-log"}

When switching to master node all works fully correct. If I run select version(); by hand on replicated DB using the same user as hasura does I got result. Looks like hasura is trying to do some kind of lock on the db while on readonly node it is not possible.

I'm using separate PostgreSQL DB as metadata database. So it looks like a regression.

What is the expected behaviour?

Database should be connected and all objects should be visible.

How to reproduce the issue?

1. Build a physically replicated PostgreSQL server
2. Enable Feature Flags:

• New Relationship tab UI for tables/views
• Enable new Table Tracking UI for Postgres & SQL Server
• Data Tab Performance Mode

3. Try to connect to replicated node

Screenshots or Screencast

Please provide any traces or logs that could help here.

Any possible solutions/workarounds you're aware of?

It looks like if the metadata will be build upon head node you can still query the slave node for data.

Keywords

cannot set transaction read-write mode during recovery