BUG #17698: On SIGTERM, psql terminates, but leaves the statement running
The following bug has been logged on the website:
Bug reference: 17698
Logged by: Dmitry Astapov
Email address: dastapov@gmail.com
PostgreSQL version: 14.2
Operating system: CentOS 7
Description:
Reproduction: start psql, launch a query that will take a while (I used
'select pg_sleep(1234)'), then from another terminal do "kill -TERM <pid of
psql>".
Psql would be terminated, but query would not be, and will keep running
until TCP timeout kicks in and kills the stale connection.
SIGINT, on the other hand, will cause psql will terminate the current
statement.
Perhaps psql could have a SIGTERM handler that does roughly what SIGINT
handler does before terminating the whole process, and would cancel the
current statement (if any)?
On 2022-Nov-25, PG Bug reporting form wrote:
Reproduction: start psql, launch a query that will take a while (I
used 'select pg_sleep(1234)'), then from another terminal do
"kill -TERM <pid of psql>". Psql would be terminated, but query would
not be, and will keep running until TCP timeout kicks in and kills the
stale connection.SIGINT, on the other hand, will cause psql will terminate the current
statement.Perhaps psql could have a SIGTERM handler that does roughly what
SIGINT handler does before terminating the whole process, and would
cancel the current statement (if any)?
Sure. I doubt we want to wait until a reply to the cancel request is
obtained, but we could have the sigterm handler at least send it before
terminating.
Why do you send SIGTERM to psql anyway?
--
Álvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/
"Thou shalt not follow the NULL pointer, for chaos and madness await
thee at its end." (2nd Commandment for C programmers)
On Mon, Jan 16, 2023 at 3:21 PM Alvaro Herrera <alvherre@alvh.no-ip.org>
wrote:
On 2022-Nov-25, PG Bug reporting form wrote:
Perhaps psql could have a SIGTERM handler that does roughly what
SIGINT handler does before terminating the whole process, and would
cancel the current statement (if any)?Sure. I doubt we want to wait until a reply to the cancel request is
obtained, but we could have the sigterm handler at least send it before
terminating.Why do you send SIGTERM to psql anyway?
The way the story was told to me by the user, they started psql from inside
the neovim session and then cancelled it with Ctrl-C. Apparently, neovim
(unlike, say, vim) sends SIGTERM to the child process in this case (whereas
vim will send SIGINT).
--
D. Astapov
po 16. 1. 2023 v 20:45 odesílatel Dmitry Astapov <dastapov@gmail.com>
napsal:
On Mon, Jan 16, 2023 at 3:21 PM Alvaro Herrera <alvherre@alvh.no-ip.org>
wrote:On 2022-Nov-25, PG Bug reporting form wrote:
Perhaps psql could have a SIGTERM handler that does roughly what
SIGINT handler does before terminating the whole process, and would
cancel the current statement (if any)?Sure. I doubt we want to wait until a reply to the cancel request is
obtained, but we could have the sigterm handler at least send it before
terminating.Why do you send SIGTERM to psql anyway?
The way the story was told to me by the user, they started psql from
inside the neovim session and then cancelled it with Ctrl-C. Apparently,
neovim (unlike, say, vim) sends SIGTERM to the child process in this case
(whereas vim will send SIGINT).
I think there may be a bug - I am not able to reproduce it, but sometimes I
can crash psql by CTRL-C too.
Regards
Pavel
Show quoted text
--
D. Astapov
On Mon Jan 16, 2023 at 9:14 AM CST, Alvaro Herrera wrote:
On 2022-Nov-25, PG Bug reporting form wrote:
Reproduction: start psql, launch a query that will take a while (I
used 'select pg_sleep(1234)'), then from another terminal do
"kill -TERM <pid of psql>". Psql would be terminated, but query would
not be, and will keep running until TCP timeout kicks in and kills the
stale connection.SIGINT, on the other hand, will cause psql will terminate the current
statement.Perhaps psql could have a SIGTERM handler that does roughly what
SIGINT handler does before terminating the whole process, and would
cancel the current statement (if any)?Sure. I doubt we want to wait until a reply to the cancel request is
obtained, but we could have the sigterm handler at least send it before
terminating.
I am currently working on a patch for this.
--
Tristan Partin
Neon (https://neon.tech)