You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Postgres distributed lock using PgBouncer connection pooler #168
my postgres database is behind a PgBouncer connection pooler with transaction pooling and it does not allow prepared statements, so its failing as the Distributed Lock postgres implementation uses prepared statement:
@madelson Is there any reason why you support pg_advisory_lock() but not pg_advisory_xact_lock()? In continuation of our discussion in #175 , I read that with pgbouncer in transaction mode this is the only way to correctly set up advisory locks for PostgreSQL.
@P4rpleSky im not opposed to adding support for that and in fact we do support the equivalent capability for SQL server.
IIRC the reason I didn’t do it initially was that in SQL server I had bad experiences with transaction-based locks. At least on that database keeping transactions open for a long time can cause issues like blocking the transaction log from being flushed. I have no idea if the same is true for Postgres (of course, not all applications need to hold distributed locks for a long time).
can you point me to the resource where you were reading about using transaction-based locks with pg_bouncer?
The problem with prepared statements is just the beginning. My comment above is related to this type of problem: if our application unexpectedly shuts down (or our database connection is closed) while having a session-based lock, we won't be sure that pgbouncer with transaction pooling won't make us a lock forever. With transaction-based locks, we'll guarantee that even in the situation described above, it will be unlocked because the transaction will end.
Maybe this problem is somehow solved by calling pg_sleep in your code, but I'm not sure about it. Can you tell me something about it, please? I will be very grateful!
Could you please provide an update on whether the issue with accessing the database via PgBouncer has been resolved? Thanks in advance for your assistance.
@tiago156@P4rpleSky@droivan I've implemented transaction-based locking in the upcoming release. Can anyone confirm that we do not need to disable prepared statements (per the original ask)?
Activity
madelson commentedon Aug 21, 2023
@tiago156 this is an interesting issue and not one I was aware of. Suporting this will required a change to the library. A few questions:
PosgresDatabaseConnectioncould detect that pgbouncer is used (e.g. by inspecting the connection string)?madelson commentedon Oct 7, 2023
@tiago156 have you looked at https://stackoverflow.com/questions/20150503/prepared-statement-doesnt-exist ? Seems like possibly this is fixable with PG configuration?
P4rpleSky commentedon Nov 22, 2023
@madelson Is there any reason why you support
pg_advisory_lock()but notpg_advisory_xact_lock()? In continuation of our discussion in #175 , I read that withpgbouncerin transaction mode this is the only way to correctly set up advisory locks for PostgreSQL.madelson commentedon Nov 22, 2023
@P4rpleSky im not opposed to adding support for that and in fact we do support the equivalent capability for SQL server.
IIRC the reason I didn’t do it initially was that in SQL server I had bad experiences with transaction-based locks. At least on that database keeping transactions open for a long time can cause issues like blocking the transaction log from being flushed. I have no idea if the same is true for Postgres (of course, not all applications need to hold distributed locks for a long time).
can you point me to the resource where you were reading about using transaction-based locks with pg_bouncer?
P4rpleSky commentedon Nov 22, 2023
@madelson Here are some useful links:
The problem with prepared statements is just the beginning. My comment above is related to this type of problem: if our application unexpectedly shuts down (or our database connection is closed) while having a session-based lock, we won't be sure that
pgbouncerwith transaction pooling won't make us a lock forever. With transaction-based locks, we'll guarantee that even in the situation described above, it will be unlocked because the transaction will end.Maybe this problem is somehow solved by calling
pg_sleepin your code, but I'm not sure about it. Can you tell me something about it, please? I will be very grateful!droivan commentedon Jan 30, 2024
Hi, @madelson, @tiago156!
Could you please provide an update on whether the issue with accessing the database via PgBouncer has been resolved? Thanks in advance for your assistance.
madelson commentedon Jan 31, 2024
@P4rpleSky thanks for the great info. I think we should add transactional pg lock support to the library.
would you or @droivan (or anyone else) be interested in contributing?
Implement transaction-scoped locks for postgres
madelson commentedon Mar 17, 2024
@tiago156 @P4rpleSky @droivan I've implemented transaction-based locking in the upcoming release. Can anyone confirm that we do not need to disable prepared statements (per the original ask)?
rafaelkallis commentedon Apr 24, 2024
@madelson I can confirm the prepared statement exception has disappeared once I used
UseTransaction().OlegUfaev commentedon Aug 2, 2024
Also, you need to change max_prepared_statements to a non-zero value in
pgbouncerconfiguration to enable Protocol-level prepared plans