Обсуждение: Timing out A Blocker Based on Time or Count of Waiters

On 3/22/24 09:25, Fred Habash wrote:
> Facing an issue where sometimes humans login to a database and run DDL 
> statements causing a long locking tree of over 1000 waiters. As a 

The above needs more explanation:

1) Define locking tree.

2) Define waiters.

3) Provide examples of the DDL.


> workaround, we asked developers to always start their DDL sessions 
> with 'SET lock_timeout = 'Xs'.
> 
> I reviewed the native lock timeout parameter in Postgres and found 7. 
> None seem to be related to blocker timeouts directly.
> 
> idle_in_transaction_session_timeout
> idle_session_timeout
> lock_timeout: How long a session waits for a lock
> statement_timeout
> authentication_timeout
> deadlock_timeout
> log_lock_waits
> 
> Instead, I put together a quick procedure that counts waiter sessions 
> for a given blocker and terminates it if waiter count exceeds a threshold.
> 
> Is there not a native way to ...
> 1. Automatically time out a blocker
> 2. A metric that shows how many waiters for a blocker?
> 
> Thanks
> -- 
> 
> ----------------------------------------
> Thank you
> 
> 

-- 
Adrian Klaver
adrian.klaver@aklaver.com

> On Mar 22, 2024, at 09:25, Fred Habash <fmhabash@gmail.com> wrote:
>
> Facing an issue where sometimes humans login to a database and run DDL statements causing a long locking tree of over
1000waiters. As a workaround, we asked developers to always start their DDL sessions with 'SET lock_timeout = 'Xs'.  
>
> I reviewed the native lock timeout parameter in Postgres and found 7. None seem to be related to blocker timeouts
directly.

"Blocker" isn't totally clear to me, but assuming you mean, "Is there a way of terminating a transaction that is
holdinga lock on which other sessions are waiting after a certain amount of time, even if the session is actively
runningqueries?", the answer is no, there's no specific setting in PostgreSQL that does that automatically. 

The most apropos setting is `idle_in_transaction_session_timeout`, but that will not terminate a session that is
activelyrunning a query.  A combination of `idle_in_transaction_session_timeout` and `statement_timeout` will get you
veryclose to it, however.  That won't catch a session that is running queries that are less take less than
`statement_timeout`to complete, and don't wait more than `idle_in_transaction_session_timeout` to issue a new query.
Itdoes not also discriminate between transactions that are holding locks on which other sessions are waiting, and ones
thataren't. 

You could write a polling script that checks pg_stat_activity and pg_locks, and terminates transactions that have been
runninglonger than x seconds, and which are holding locks that other sessions are waiting on.  I'm not sure that's
reallya recommended course of action, though, as scripts like that can often kill things you didn't really mean them
to.

On 3/22/24 12:41, Fred Habash wrote:
> Lock tree: All PID's waiting on a lock held by/blocked by single blocker 
> PID. Similar to what you see in the output of this script: 
> https://github.com/dataegret/pg-utils/blob/master/sql/locktree.sql 
> <https://github.com/dataegret/pg-utils/blob/master/sql/locktree.sql> . 
> It uses the dot connotation to draw a tree.
> 
> Waiters: The PID (first column) returned by this query, for example
> 
> SELECT
>      activity.pid,
>      activity.usename,
>      activity.query,
>      blocking.pid AS blocking_id,
>      blocking.query AS blocking_query
> FROM pg_stat_activity AS activity
> JOIN pg_stat_activity AS blocking ON blocking.pid = 
> ANY(pg_blocking_pids(activity.pid));
> 
> DDL example: An 'alter table ... alter column ...' would cause all DML 
> and SELECT statements to wait/block.
> 
> Hope this answers your question. Thanks for your interest.

Yes.

To me the issue is,

"Facing an issue where sometimes humans login to a database and run DDL 
statements causing a long locking tree of over 1000 waiters."

where the problem is people running disruptive statements without regard 
to planning or what else is happening on the database. I am not sure 
that dropping a statement just based on a count is progress. If the DDL 
is important then it needs to be run at some point and you are 
conceivably back at the same blocking issue. This then leads to two 
possibilities either the DDL is not important and shouldn't be run or it 
is and some thought and timing needs to be applied before it is run.

> 
> On Fri, Mar 22, 2024 at 12:32 PM Adrian Klaver 
> <adrian.klaver@aklaver.com <mailto:adrian.klaver@aklaver.com>> wrote:
> 
>     On 3/22/24 09:25, Fred Habash wrote:
>      > Facing an issue where sometimes humans login to a database and
>     run DDL
>      > statements causing a long locking tree of over 1000 waiters. As a
> 
>     The above needs more explanation:
> 
>     1) Define locking tree.
> 
>     2) Define waiters.
> 
>     3) Provide examples of the DDL.
> 
> 
>      > workaround, we asked developers to always start their DDL sessions
>      > with 'SET lock_timeout = 'Xs'.
>      >
>      > I reviewed the native lock timeout parameter in Postgres and
>     found 7.
>      > None seem to be related to blocker timeouts directly.
>      >
>      > idle_in_transaction_session_timeout
>      > idle_session_timeout
>      > lock_timeout: How long a session waits for a lock
>      > statement_timeout
>      > authentication_timeout
>      > deadlock_timeout
>      > log_lock_waits
>      >
>      > Instead, I put together a quick procedure that counts waiter
>     sessions
>      > for a given blocker and terminates it if waiter count exceeds a
>     threshold.
>      >
>      > Is there not a native way to ...
>      > 1. Automatically time out a blocker
>      > 2. A metric that shows how many waiters for a blocker?
>      >
>      > Thanks
>      > --
>      >
>      > ----------------------------------------
>      > Thank you
>      >
>      >
> 
>     -- 
>     Adrian Klaver
>     adrian.klaver@aklaver.com <mailto:adrian.klaver@aklaver.com>
> 
> 
> 
> -- 
> 
> ----------------------------------------
> Thank you
> 
> 

-- 
Adrian Klaver
adrian.klaver@aklaver.com