Use Primary DB when conection to read replica is refused.

[LaurinBrechter]

I am using drizzle with a primary cloud db and a read replica that is co-located with my server. I am doing something like this to set up the connection:

export const masterDb = drizzle({
  connection: process.env.DATABASE_URL!,
  schema: { ...schema, ...authSchema },
  logger: new MyLogger("master"),
});

export const replicaDb = drizzle({
  connection: process.env.DATABASE_URL_REPLICA ?? process.env.DATABASE_URL!,
  schema: { ...schema, ...authSchema },
  logger: new MyLogger("replica"),
});

export const db = withReplicas(masterDb, [replicaDb]);

Now, the problem is that when my local db goes down I get an error for read operations. This would be manageable for queries that I make myself since I could make a wraper function that checks for an ECONNREFUSED error when making the query and then switches to the primary. But there's external libraries like better-auth or next-auth to which you can only pass the db object.

ERROR [Better Auth]: INTERNAL_SERVER_ERROR Error: connect ECONNREFUSED 127.0.0.1:5632
    at <unknown> (Error: connect ECONNREFUSED 127.0.0.1:5632)
    at async authCached (src/lib/authUtils.ts:87:15)
    at async Header (src/components/header/Header.tsx:16:15)
  85 |   }
  86 |
> 87 |   const data = await auth.api.getSession({
     |               ^
  88 |     headers: await headers(),
  89 |   });
  90 | {
  errno: -111,
  code: 'ECONNREFUSED',
  syscall: 'connect',
  address: '127.0.0.1',
  port: 5632
}

Is there an option for drizzle to detect that the read replica is down and then switch to the primary? Another option I see is that one checks in the callback that decides which db to use if the read replica is still there:

const db = withReplicas(primaryDb, [read1, read2], (replicas) => {
     // try connecting to the read replica and if that fails only return primary
});

[jericopulvera]

I encountered this issue as well. I just followed the documentation. and there's no automatic fallback to the primary. When my replica hits its connection limit, it just throws an error. It's not ideal.

[SooditK]

@AndriiSherman (sorry to tag you directly 🙇🏼) but can we please have this? atleast a documentation update would be cool

[SaulMoreno3]

withReplicas is a replica-selection mechanism, not a health-check or automatic failover layer. If the selected replica connection returns ECONNREFUSED, the query rejects; Drizzle does not transparently retry that read against the primary.

The callback can be useful for choosing a replica based on application state, but it should not perform a network probe for every query. For libraries such as Better Auth that accept only one db object, the usual options are:

• pass the primary connection for correctness-critical/authentication reads;
• put failover in the database proxy or managed database layer;
• wrap the read path in a small adapter that retries only idempotent reads after a connection error, then falls back to the primary.

Keep the retry bounded and do not apply it blindly to writes or transactions, because a connection error does not always tell you whether the server received the previous operation.