Postgres Not Starting: The Case of the Missing Output
Strangely quiet around here
My colleague Bob Pacheco asked me to help with a strange problem he was witnessing for a client. A new Postgres cluster was getting created on a Kubernetes node, but it refused to start. More mysteriously, there was no message in the Postgres logs, nothing in the pg_ctl start up file, and no output anywhere. The pg_ctl program started up, then ended. Nothing on stderr, nothing on stdout! We were able to duplicate it on the command line using pg_ctl. Oddly enough, a syntax error in the postgresql.conf would cause output (an error). That part happens very early, so something was happening after that point - but before the logging collector started up. Time to go to the lower level.
Enter the postmaster
The executable program that actually runs Postgres itself is called, aptly enough, postgres. You may still see it referred to as postmaster. Before version 16 of Postgres, there was a symlink named postmaster that pointed to the postgres executable. Even though that symlink is no longer created, you will still see this name in use, such as the postmaster.pid
file that appears in the data directory when you start up Postgres.
Tools like pg_ctl and systemd call this postgres executable itself, making sure to pass in parameters, set up stderr redirects, and so forth. But you can run it yourself directly! This is a great debugging tool to see what is happening as Postgres comes to life. For this session, we decided to start it in single
mode, which creates a "stand-alone backend". In other words, it brings Postgres up, but does not let anyone else in, does not start any background processes (e.g. autovacuum), and does not listen for incoming connections. In short, a nice and simplified Postgres to flush out problems. To use it, run the executable, provide --single
as the first argument, and tell it where the data directory lives.
Our first attempt looked like this and immediately revealed a bus error and a core dump!
$ bin/postgres --single -D data
Bus Error(Core Dumped)
Well, that explains the lack of output! A Bus Error occurs when a process is trying to access memory it is not allowed to. The kernel denies the attempt and sends a SIGBUS signal. The core dump indicates that the postgres executable was caught off guard by this, and crashed before it could do any sort of cleanup or output any messages. Hence, the lack of output that started this whole process.
Narrowing down the problem
One trick to rule out any settings or extensions is to load Postgres in a pristine state - in other words, with all configuration values at their default. To do that, you can provide your own configuration file for the postgres executable to read. Creating an empty file will do the trick:
$ touch empty
$ bin/postgres --single -D data --config-file=empty
Bus Error(Core Dumped)
Still no change, so we can rule out extensions and configuration items. The next step was to run it with debugging cranked all the way up. The postgres executable takes a -d
argument, with values from 1 to 5, so we ran:
$ bin/postgres --single -D data -d 5
2023-12-25 12:34:00.767 EST [1989] DEBUG: invoking IpcMemoryCreate(size=1515356162)
Bus Error(Core Dumped)
The "Bus Error" was a strong clue that this was memory related, and we can see that the last normal thing it did was request some shared memory from the operating system. Normally, I would blame SELinux for mysterious start up problems. However, if memory is involved, there is a better suspect to consider: huge pages. In the past, Linux would allocate physical memory in blocks of 4KB. As most modern programs require a lot of memory, that meant gathering up and doing housekeeping for a LOT of those 4KB chunks. With huge pages, the kernel can use much, much larger blocks. You can see your current size with:
grep Hugepagesize /proc/meminfo
Fixing a huge problem
The next step was to try and load Postgres without using huge pages. The default value for current versions of Postgres is huge_pages=try
so we wanted to force it to "off". By passing a configuration item directly on the command line, we can ensure that it is forced off, as command-line arguments override anything set in the configuration files.
$ bin/postgres --single -D data --huge_pages=off
2023-12-25 12:35:00.767 EST [1991] DEBUG: invoking IpcMemoryCreate(size=1515356162)
2023-12-25 12:35:54.961 EST [1991] DEBUG: dynamic shared memory system will support 1674 segments
2023-12-25 12:35:54.961 EST [1991] DEBUG: created dynamic shared memory control segment 2743578136 (269765 bytes)
2023-12-25 12:35:54.966 EST [1991] DEBUG: InitPostgres
2023-12-25 12:35:54.966 EST [1991] DEBUG: my backend ID is 1
...
Voila - our database started up! After we made the huge_pages change permanent in the postgresql.conf file, the cluster started up without any problem. But what exactly was going on with huge pages here?
It turns out this was a linux kernel bug, in which the kernel was happy to pass you the memory to use, but then got upset when you tried to actually use it. Obviously, it should have denied the initial request, a situation Postgres is well-equipped to handle gracefully, without any core dumps. I found this thread on the Linux kernel list that would seem to be the bug in question. We also verified that the 4.18 kernel version we were using was before the fixed version in 4.18.0.513. So presumably once the kernel is newer, we can switch huge_pages
back to "try"!
See more about our work on fixing this linux kernel issue with Crunchy Postgres for Kubernetes.
Lessons learned
- When troubleshooting, start with complex then go down to simple. In this case, we used the
postgres
executable directly. - All the parts of your stack are important. This was a Linux kernel bug, but other systems could affect Postgres as well.
- Keep everything up to date. In this case, a newer version of the kernel had already fixed the problem.
- Huge pages are overall a good thing. Leave it as "try" unless you can prove it is causing problems.
Related Articles
- Crunchy Data Warehouse: Postgres with Iceberg for High Performance Analytics
8 min read
- Loading the World! OpenStreetMap Import In Under 4 Hours
6 min read
- Easy Totals and Subtotals in Postgres with Rollup and Cube
5 min read
- A change to ResultRelInfo - A Near Miss with Postgres 17.1
8 min read
- Accessing Large Language Models from PostgreSQL
5 min read