Re: Enhance pg_dump multi-threaded streaming (WAS: Re: filesystem full during vacuum - space recovery issues)

Hi Andrew,

This is very interesting.

I had started looking at pg_dumpall trying to work out an approach.  I noticed parallel.c essentially already does all the thread creation and coordination that I knew would be needed.  Given that is a solved problem, I started to look further (continued below).

On 2024-07-19 Fr 9:46 AM, Thomas Simpson wrote:

Hi Scott,

I realize some of the background was snipped on what I sent to the hacker list, I'll try to fill in the details.

Short background is very large database ran out of space during vacuum full taking down the server.  There is a replica which was applying the WALs and so it too ran out of space.  On restart after clearing some space, the database came back up but left over the in-progress rebuild files.  I've cleared that replica and am using it as my rebuild target just now.

Trying to identify the 'orphan' files and move them away always led to the database spotting the supposedly unused files having gone and refusing to start, so I had no successful way to clean up and get space back.

Last resort after discussion is pg_dumpall & reload.  I'm doing this via a network pipe (netcat) as I do not have the vast amount of storage necessary for the dump file to be stored (in any format).

On 19-Jul-2024 09:26, Scott Ribe wrote:

Do you actually have 100G networking between the nodes? Because if not, a single CPU should be able to saturate 10G.

Servers connect via 10G WAN; sending is not the issue, it's application of the incoming stream on the destination which is bottlenecked.

Likewise the receiving end would need disk capable of keeping up. Which brings up the question, why not write to disk, but directly to the destination rather than write locally then copy?

In this case, it's not a local write, it's piped via netcat.

Do you require dump-reload because of suspected corruption? That's a tough one. But if not, if the goal is just to get up and running on a new server, why not pg_basebackup, streaming replica, promote? That depends on the level of data modification activity being low enough that pg_basebackup can keep up with WAL as it's generated and apply it faster than new WAL comes in, but given that your server is currently keeping up with writing that much WAL and flushing that many changes, seems likely it would keep up as long as the network connection is fast enough. Anyway, in that scenario, you don't need to care how long pg_basebackup takes.

If you do need a dump/reload because of suspected corruption, the only thing I can think of is something like doing it a table at a time--partitioning would help here, if practical.

The basebackup is, to the best of my understanding, essentially just copying the database files.  Since the failed vacuum has left extra files, my expectation is these too would be copied, leaving me in the same position I started in.  If I'm wrong, please tell me as that would be vastly quicker - it is how I originally set up the replica and it took only a few hours on the 10G link.

The inability to get a clean start if I move any files out the way leads me to be concerned for some underlying corruption/issue and the recommendation earlier in the discussion was opt for dump/reload as the fail-safe.

Resigned to my fate, my thoughts were to see if there is a way to improve the dump-reload approach for the future.  Since dump-reload is the ultimate upgrade suggestion in the documentation, it seems worthwhile to see if there is a way to improve the performance of that especially as very large databases like mine are a thing with PostgreSQL.  From a quick review of pg_dump.c (I'm no expert on it obviously), it feels like it's already doing most of what needs done and the addition is some sort of multi-thread coordination with a restore client to ensure each thread can successfully complete each task it has before accepting more work.  I realize that's actually difficult to implement.

There is a plan for a non-text mode for pg_dumpall. I have started work on it, and hope to have a WIP patch in a month or so. It's not my intention to parallelize it for the first cut, but it could definitely be parallelizable in future. However, it will require writing to disk somewhere, albeit that the data will be compressed. It's well nigh impossible to parallelize text format dumps.

Restoration of custom and directory format dumps has long been parallelized. Parallel dumps require directory format, and so will non-text pg_dumpall.

My general approach (which I'm sure is naive) was:

Add to pg_dumpall the concept of backup phase and I have the basic hooks in place.  0 = role grants etc.  The stuff before dumping actual databases.  I intercepted the fprintf(OPF to a hook function that for normal run just ends up doing the same as fprintf but for my parallel mode, it has a hook to send the info via the network (still to be done but I think I may need to alter the fprintf stuff with more granularity of what is being processed at each output to help this part, such as outputRoleCreate, outputComment etc.).

Each subsequent phase is a whole database - increment at each pg_dump call.  The actual pg_dump is to get a new format, -F N for network; based around directory dump as the base, my intention was to make multiple network pipes to send the data in place of the files within the directory.  Essentially relying on whatever is already done to organize parallel dumps to disk to be sufficient for coordinating network streaming.

The restore side needs to do network listen plus some handshaking to confirm completion of the incoming phases, any necessary dependency tracking on restore etc.

My goal was to actively avoid the disk usage part through the coordination over the network between dump and restore even though my starting point is the pg_backup_directory code.  Any problem on the restore side would feed back and halt the dump side in error so this is a new failure mode compared with how it works just now.

I'll hold off a bit as I'm very interested in any feedback you have, particularly if you see serious flaws in my though process here.

cheers


andrew



--
Andrew Dunstan
EDB: https://www.enterprisedb.com

Thanks

Tom