On 2020-05-19 09:22, Anders Bøgh Bruun wrote:
I have run into a (to me) weird issue with logical replication. We are running Zalandos postgres-operator in our Kubernetes clusters and have recently had a use-case where we wanted to start doing logical replication of select tables to a data warehouse, also running postgres. It worked as expected at first, but then after a pod-restart in Kubernetes, the replication slots that were created for the subscription were gone. A bit of reading later, and I learn we need to tell Patroni which slots should be permanently available, so we specify a slot and try to set this up, but then run into an error which says the publication does not exist, even though we can verify that it does. At first I suspected Patroni handling the replication slots to be the cause of the problem, but about a week's worth of learning and experimenting later, I can now reliably replicate the problem in pure postgres. Patroni is kind of the catalyst, since my findings are that if the replication slot is created before data is inserted into the source database, and a publication is created, then it breaks. If the replication slot is created after data is inserted and the publication is created, then it works. We just can't tell Patroni to not create it until some arbitrary point in time. I am guessing this is either a bug or a case of us not knowing what we are doing...
What does "works" and "breaks" mean? Are there error messages shown from commands or in the server logs? Is replication progressing, but doing something your are not expecting? etc.
-- Peter Eisentraut http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
