Rollup node publishes commitments with a large delay and not being able to catch up

I observe a strange behavior of the injector of a rollup node:

• It does not publish commitments once in a few blocks as usual
• Rather it attempts to publish them in batches every few hours
• Most times without luck, but no errors in the logs
• At a certain point it reaches Stack overflow, restart helps

I use this Octez branch: !8785 (merged) Rollup deployment: https://mumbainet.tzkt.io/sr1TKmtdm9XaFFHe7Zfx7zZXaK6hkp2z6qLa/operations/

How to reproduce (need to specify your secret key with a non-zero balance in Mumbainet):

docker run --name tz-rollup-operator -d -v rollup-node:/root/.tezos-smart-rollup-node -p 127.0.0.1:8932:8932 -e ROLLUP_ADDRESS=sr1TKmtdm9XaFFHe7Zfx7zZXaK6hkp2z6qLa -e OPERATOR_KEY=<edsk...> -e NETWORK=mumbainet -e NODE_URI=https://rpc.tzkt.io/mumbainet ghcr.io/baking-bad/tz-rollup-operator:pr-27 rollup-node