FYI only; this does not need an answer.
I was working on a Postgres server under heavy load, and issued a GRANT command that hung. It was not blocked by any other commands.
I had a few open connections and was able to kill several of the processes with a normal pg_cancel_backend (SIGTERM) command, but my GRANT command didn't respond to either that or pg_terminate_backend (SIGINT). I finally tried "kill -9 (pid)" (SIGKILL) and the server crashed.
Issuing SIGKILL to the database server process or the postmaster can cause crashes--that's well documented. Running SIGKILL against a child process can also crash the database.
Running SIGKILL against a child process can also crash the database
Any fatal signal that terminates any backend without a chance to clean up, such as SIGSEGV, SIGABRT, SIGKILL, etc, will cause the postmaster to assume that shared memory may be corrupt. It will roll back all transactions, terminate all running backends, and restart.
PostgreSQL does that to protect your data. If something went wrong before a backend crashed that caused it to scribble on shared memory, then shared_buffers could contain invalid data that'd get flushed to disk and replace good pages.
I was pretty sure that was in the docs, but all I can find is what I think you were referring to in shutting down the server.
Anyway, if you SIGKILL a backend you'll see something like:
WARNING: terminating connection because of crash of another server process
DETAIL: The postmaster has commanded this server process to roll back the
current transaction and exit, because another server process exited
abnormally and possibly corrupted shared memory.
HINT: In a moment you should be able to reconnect to the database and
repeat your command.
server closed the connection unexpectedly
This probably means the server terminated abnormally
before or while processing the request.
The connection to the server was lost. Attempting reset: Succeeded.
This also happens if the OOM killer kills a backend, which is why you should turn off memory overcommit on Linux.
I wrote some guidance on things to do and not to do with PostgreSQL on my blog. Worth a look.
Related
I'm seeing my client's Cloud SQL database go into recovery mode every day or two. All connections are rejected with the error:
Every time this is proceeded by log messages stating:
server process (PID 2883588) was terminated by signal 9: Killed
The SQL query is logged, always the same query
terminating any other active server processes
Then many: terminating connection because of crash of another server process
At which point the database comes back in recovery mode for a few seconds, before continuing life as normal.
Why would Cloud SQL be terminating a query for me? I've run the query in question (on the same server) and it completes happily in under a millisecond. Could this be caused by transient server load?
Everytime i run the same query i'm getting this error:
DETAIL: The postmaster has commanded this server process to roll back the current transaction and exit, because another server process exited abnormally and possibly corrupted shared memory.
HINT: In a moment you should be able to reconnect to the database and repeat your command.
server closed the connection unexpectedly
This probably means the server terminated abnormally
before or while processing the request.
connection to server was lost
I'm using phpPgAdmin 5.1.
A concurrent query in a different database session has crashed its server backend process. As a consequence, the whole database stops and performs crash recovery from the latest checkpoint.
You should look into the database server log to see what the problem is.
I know database indexes can become corrupted if the server crashes. If I do:
sudo service postgresql restart
can that cause the same kind of corruption as a server crash?
That depends on the system I belive. You should look into the script to check the actual command issued. Eg. here we see, that restart is equal to stop & start. then checking stop we see it does killproc postmaster and removes pid. From the man killproc sends SIGTERM if otherly not specified. By the documentation
SIGTERM
This is the Smart Shutdown mode. After receiving SIGTERM, the
server disallows new connections, but lets existing sessions end their
work normally. It shuts down only after all of the sessions terminate.
If the server is in online backup mode, it additionally waits until
online backup mode is no longer active. While backup mode is active,
new connections will still be allowed, but only to superusers (this
exception allows a superuser to connect to terminate online backup
mode). If the server is in recovery when a smart shutdown is
requested, recovery and streaming replication will be stopped only
after all regular sessions have terminated.
So in presented case, indexes should survive. But you definetely should watch your /etc/init.d/ script to be sure.
I was debugging a PostgreSQL 9.2 database corruption issue (on Solaris, but I doubt it matters) recently, and I found that we could reproduce it reliably if the client died in the middle of a transaction and then I shut down PostgreSQL by doing pkill postgres (which basically sends SIGTERM to every running postgres process). If instead we did pkill -QUIT postgres to send SIGQUIT, the database would shut down cleanly and no corruption would occur.
Based on the PostgreSQL 9.2 docs, I think that SIGTERM should be 100% expected by the database server, so why is it not safe to shut down like this? Is it a bug in PostgreSQL, or could I be doing something (configuration, etc.) that would allow the corruption to occur?
I don't think sigterm is what is causing your problem. Again, recommend you ask on dba.stackexchange instead.
If the client dies in the middle of a transacction, then the problem is that the network connection hangs? And then when you kill it you get corruption during WAL replay?
This is a complicated area to troubleshoot but here are some places to begin:
What is going on conncurrently when this happens? What sort of transaction commit load?
How often do WAL logs normally get rotated?
It is possible you could be running into a rare, obscure bug with PostgreSQL (possibly somewhere between the db, kernel, and filesystem), but if so please start by upgrading to latest 9.2 and try to reproduce again. Term and even kill signals are supposed to be 100% safe on PostgreSQL so if you are seeing database corruption, that is not expected.
I am performing a bulk copy into postgres with about 80GB of data.
\copy my_table FROM '/path/csv_file.csv' csv DELIMITER ','
Before the transaction is committed I get the following error.
Server closed the connection unexpectedly
This probably means the server terminated abnormally
before or while processing the request.
In the PostgreSQL logs:
LOG:server process (PID 21122) was terminated by signal 9: Killed
LOG:terminating any other active server processes
WARNING:terminating connection because of crash of another server process
DETAIL:The postmaster has commanded this server process to roll back the current transaction and exit, because another server process exited abnormally and possibly corrupted shared memory.
HINT: In a moment you should be able to reconnect to the database and repeat your command.
Your backend process receiving a signal 9 (SIGKILL). This can happen if:
Somebody sends a kill -9 manually;
A cron job is set up to send kill -9 under some circumstances (very unsafe, do not do this); or
the Linux out-of-memory (OOM) killer triggers and terminates the process.
In the latter case you will see reports of OOM killer activity in the kernel's dmesg output. I expect this is what you'll see in your case.
PostgreSQL servers should be configured without virtual memory overcommit so that the OOM killer does not run and PostgreSQL can handle out-of-memory conditions its self. See the PostgreSQL documentation on Linux memory overcommit.
The separate question "why is this using so much memory" remains. Answering that requires more knowledge of your setup: how much RAM the server has, how much of it is free, what your work_mem and maintenance_work_mem settings are, etc. It isn't a very interesting problem to look into until you upgrade to the current PostgreSQL 8.4 patch release to make sure the problem isn't one that's already fixed.