postgresql when it's not your job

3 March 2023


Upcoming Live Presentations

I’m currently scheduled to speak at:

I hope to see you at one of these!

28 February 2023


Workers of the World, Unite!

Over the course of the last few versions, PostgreSQL has introduces all kinds of background worker processes, including workers to do various kinds of things in parallel. There are enough now that it’s getting kind of confusing. Let’s sort them all out.

You can think of each setting as creating a pool of potential workers. Each setting draws its workers from a “parent” pool. We can visualize this as a Venn diagram:

max_worker_processes sets the overall size of the worker process pool. You can never have more than that many background worker processes in the system at once. This only applies to background workers, not the main backend processes that handle connections, or the various background processes (autovacuum daemon, WAL writer, etc.) that PostgreSQL uses for its own operations.

From that pool, you can create up to max_parallel_workers parallel execution worker processes. These come in two types:

So, what shall we set these to?

Background workers that are not parallel workers are not common in PostgreSQL at the moment, with one notable exception: logical replication workers. The maximum number of these are set by the parameter max_logical_replication_workers. What to set that parameter to is a subject for another post. I recommend starting the tuning with max_parallel_workers, since that’s going to be the majority of worker processes going at any one time. A good starting value is 2-3 times the number of cores in the server running PostgreSQL. If there are a lot of cores (32 to 64 or more), 1.5 times might be more appropriate.

For max_worker_processes, a good place to start is to sum:

Then, consider max_parallel_workers_per_gather. If you routinely processes large result sets, increasing it from the default of 2 to 4-6 is reasonable. Don’t go crazy here; a query rapidly reaches a point of diminishing returns in spinning up new parallel workers.

For max_parallel_maintenance_workers, 4-6 is also a good value. Go with 6 if you have a lot of cores, 4 if you have more than eight cores, and 2 otherwise.

Remember that every worker in parallel query execution can individually consume up to work_mem in working memory. Set that appropriately for the total number of workers that might be running at any one time. Note that it’s not just work_mem x max_parallel_workers_per_gather! Each individual worker can use more than work_mem if it has multiple operations that require it, and any non-parallel queries can do so as well.

Finally, max_parallel_workers, max_parallel_maintenance_workers, and max_parallel_workers_per_gather can be set for an individual session (or role, etc.), so if you are going to run an operation that will benefit from a large number of parallel workers, you can increase it for just that query. Note that the overall pool is still limited by max_worker_processes, and changing that requires a server restart.

21 February 2023



Normally, when you drop a column from PostgreSQL, it doesn’t have to do anything to the data in the table. It just marks the column as no longer alive in the system catalogs, and gets on with business.

There is, however, a big exception to this: ALTER TABLE … SET WITHOUT OIDS. This pops up when using pg_upgrade to upgrade a database to a version of PostgreSQL that doesn’t support table OIDs (if you don’t know what and why user tables in PostgreSQL had OIDs, that’s a topic for a different time).

ALTER TABLE … SET WITHOUT OIDS rewrites the whole table, and reindexes the table as well. This can take up quite a bit of secondary storage space:

  1. On the tablespace that the current table lives in, it can take up to the size of the table as it rewrites the table.
  2. On temporary file storage (pg_tmp), it can take significant storage doing the reindexing, since it may need to spill the required sorts to disk. This can be mitigated by increasing maintenance_work_mem.

So, plan for some extended table locking if you do this. If you have a very large database to upgrade, and it still has tables with OIDs, this may be an opportunity to upgrade via logical replication rather than pg_upgrade.

16 February 2023


UUIDs vs serials for keys

This topic pops up very frequently: “Should we use UUIDs or bigints as primary keys?”

One of the reasons that the question gets so many conflicting answers is that there are really two different questions being asked:

Let’s take them independently.

Should our keys be random or sequential?

There are strong reasons for either one. The case for random keys is:

The case for sequential keys is:

Should our keys be 64 bits, or larger?

It’s often just taken for granted than when we say “random” keys, we mean “UUIDs”, but there’s nothing intrinsic about bigint keys that means they have to be sequential, or (as we noted above) about UUID keys that require they be purely random.

bigint values will be more performant in PostgreSQL than 128 bit values. Of course, one reason is just that PostgreSQL has to move twice as much data (and store twice as much data on disk). A more subtle reason is the internal storage model PostgreSQL uses for values. The Datum type that represents a single value is the “natural” word length of the processor (64 bits on a 64 bit processor). If the value fits in 64 bits, the Datum is just the value. If it’s larger than 64 bits, the Datum is a pointer to the value. Since UUIDs are 128 bits, this adds a level of indirection and memory management to handling one internally. How big is this performance issue? Not large, but it’s not zero, either.

So, if you don’t think you need 128 bits of randomness (really, 124 bits plus a type field) that a UUID provides, consider using a 64 bit value even if it is random, or if it is (for example) 16 bits of sequence plus 48 bits of randomness.

Other considerations about sequential keys

If you are particularly concerned about exposing information, one consideration is that keys that have sequential properties, even just in the high bits, can expose the rate of growth of a table and the total size of it. This may be something you don’t want run the risk of leaking; a new social media network probably doesn’t want the outside world keeping close track of the size of the user table. Purely random keys avoid this, and may be a good choice if the key is exposed to the public in an API or URL. Limiting the number of high-order sequential bits can also mitigate this, and a (probably small) cost in locality for B-tree indexes.

15 February 2023


“Database Patterns and How to Find Them” at SCaLE 2023

I’ll be speaking on Database Antipatterns and How to Find Them at SCaLE 2023, March 9-12, 2023 in Pasadena, CA.


“Extreme PostgreSQL” at PgDay/MED

I’m very happy that I’ll be presenting “Extreme PostgreSQL” at PgDay/MED in Malta (yay, Malta!) on 13 April 2023.

8 February 2023


Xtreme PostgreSQL!

The slides from my talk at the February 2023 SFPUG Meeting are now available.

30 January 2023


Nordic PgDay 2023

I’m very pleased to be talking about real-life logical replication at Nordic PgDay 2023, in beautiful Stockholm.

18 January 2023


A foreign key pathology to avoid

There’s a particular anti-pattern in database design that PostgreSQL handles… not very well.

For example, let’s say you are building something like Twitch. (The real Twitch doesn’t work this way! At least, not as far as I know!) So, you have streams, and you have users, and users watch streams. So, let’s do a schema!

CREATE TABLE stream (stream_id bigint PRIMARY KEY);

CREATE TABLE "user" (user_id bigint PRIMARY KEY);

CREATE TABLE stream_viewer (    
    stream_id bigint REFERENCES stream(stream_id),    
    user_id bigint REFERENCES "user"(user_id),   
    PRIMARY KEY (stream_id, user_id));

OK, schema complete! Not bad for a day’s work. (Note the double quotes around "user". USER is a keyword in PostgreSQL, so we have to put it in double quotes to use as a table name. This is not great practice, but more about double quotes some other time.)

Let’s say we persuade a very popular streamer over to our platform. They go on-line, and all 1,252,136 of our users simultaneously log on and start following that stream.

So, we now have to insert 1,252,136 new records into stream_viewer. That’s pretty bad. But what’s worse is now we have 1,252,136 records with a foreign key relationship to a single record in stream. During the operation of the INSERT statement, the transaction that is doing the INSERT will take a FOR KEY SHARE lock on that record. This means that at any one moment, several thousand different transactions will have a FOR KEY SHARE lock on that record.

This is very bad.

If more than one transaction at a time has a lock on a single record, the MultiXact system handles this. MultiXact puts a special transaction ID in the record that’s locked, and then builds an external data structure that holds all of the transaction IDs that have locked the record. This works great… up to a certain size. But that data structure is of fixed size, and when it fills up, it spills onto secondary storage.

As you might imagine, that’s slow. You can see this with lots of sessions suddenly waiting on various MultiXact* lightweight locks.

You can get around this in a few ways:

Not many systems have this particular design issue. (You would never actually build a streaming site using that schema, just to start.) But if you do, this particular behavior is a good thing to avoid.

16 January 2023


OK, sometimes you can lock tables.

Previously, I wrote that you should never lock tables. And you usually shouldn’t! But sometimes, there’s a good reason to. Here’s one.

When you are doing a schema-modifying operation, like adding a column to a table, PostgreSQL needs to take an ACCESS EXCLUSIVE lock on the table while it is modifying the system catalogs. Unless it needs to rewrite the table, this lock isn’t held for very long.

However, locks in PostgreSQL are first-come, first-served. If the system is busy, there may be conflicting locks on the table that you are attempting to modify. (Even just a SELECT statement takes lock on the tables it is operating on; it just doesn’t conflict with much.) If the ALTER TABLE statement can’t get the lock right away, it enters a queue, waiting to get to the front and get the lock.

However, now, every lock after that enters the queue, too, behind that ALTER TABLE. This can create the result of a long-running ACCESS EXCLUSIVE lock, even though it’s not granted. On a busy table on a busy system, this can shut things down.

So, what to do?

You can do this:

DO $$
   FOR i IN 1 .. 1000 LOOP
      EXCEPTION WHEN lock_not_available THEN
         PERFORM pg_sleep(1);
   RAISE lock_not_available;

This loops until it can acquire the lock, but doesn’t sit in the queue if it can’t. Once it acquires the lock, it does the modification and exits. If it can’t acquire the lock after a certain number of cycles, it exits with an error (you can set the number of cycles to anything, and you can adjust time it sleeps after failing to get the lock).

« Older Entries

Newer Entries »