Arktekk.no - Blog

Robin Kåveland

Published 2024-04-12

postgres

It is rewarding to work on software that people care about and use all around the clock. This constant usage means we can't simply take the system offline for maintenance without upsetting users. Therefore, techniques that allow us to update the software seamlessly without downtime or compromising service quality are incredibly valuable.

Most projects I've worked on use a relational database for persistence, and have some sort of migration tool like flyway or liquibase to make changes to the database schema. This post is about a particular kind of migration situation that, in my experience, most developers who work on such projects will encounter at some point in their career. They will want to apply a simple, and seemingly innocent migration, like adding a column to a table and it'll cause some number of requests to fail, or maybe even a small outage. There are some tricks we can use here to reduce risk and automatically detect some patterns that cause this problem.

This post was discussed on hacker news here.

Scenario 1

We'll use postgres to describe the setup and scenario, with the following simple schema:

CREATE TABLE documents(id serial PRIMARY KEY, document text);

This scenario has 2 transactions in it:

A is a transaction that wants to complete 2 statements, one fast and one slow, both modifications to the database.
B is a transaction that only wants to read a single document.

I will start these in different psql shells and we'll take a look at what happens.

`A` adds a column to `documents`

postgres=# BEGIN;
BEGIN
postgres=*# ALTER TABLE documents ADD COLUMN is_indexed bool;
ALTER TABLE

A is now holding an AccessExclusiveLock on documents, and all subsequent transactions that attempt to do anything with that table must wait:

postgres=# SELECT 
  l.relation::regclass AS object_name,
  l.mode, 
  locktype, 
  granted
FROM pg_locks l
JOIN pg_class c ON c.oid = l.relation
JOIN pg_namespace n ON n.oid = c.relnamespace
AND n.nspname != 'pg_catalog';
 object_name |        mode         | locktype | granted
-------------+---------------------+----------+---------
 documents   | AccessExclusiveLock | relation | t
(1 row)

`A` starts a slow statement

Let's say, then that A starts a slow statement, like filling in all the missing values for the new column (suppose that documents is a big table):

postgres=*# UPDATE documents SET is_indexed = true;

`B` attempts to read a single document

postgres=# BEGIN;
BEGIN
postgres=*# SELECT document FROM documents WHERE id = 3;

B is now blocked and must wait for A to finish:

# SELECT 
  l.relation::regclass AS object_name, 
  l.mode, 
  locktype, 
  granted
FROM pg_locks l
JOIN pg_class c ON c.oid = l.relation
JOIN pg_namespace n ON n.oid = c.relnamespace
AND n.nspname != 'pg_catalog';
  object_name   |        mode         | locktype | granted
----------------+---------------------+----------+---------
 documents_pkey | RowExclusiveLock    | relation | t
 documents      | RowExclusiveLock    | relation | t
 documents      | AccessExclusiveLock | relation | t
 documents      | AccessShareLock     | relation | f
(4 rows)

Once A commits successfully, B immediately executes. In this scenario, if B is a very frequent transaction type and A is very slow, this can seem like a partial outage that can impact users on the live system.

This scenario is essentially identical to one in which there's a single migration statement that requires an AccessExclusiveLock and performs a table rewrite, such as adding a NOT NULL column with a DEFAULT value*.

But at least if A is fast, nothing bad can happen, right? Well...

Scenario 2

This scenario has 3 transactions in it:

A is a slow transaction. It reads documents and puts all the documents somewhere, possibly a search engine or something. The search engine is having a particularly bad day today or perhaps A is just reading more data than it should.
B is our migration. It wants to add a column to documents, so that the job that runs A can see which documents that have already been sent to the search engine.
C represents one of many fast transactions, needle-in-the-haystack kind of transactions. For example users that have clicked a link from the search engine and want to retrieve the entire document to read it, instead of the summary from the search engine.

There's a particular interleaving of these that could cause a lot of C transactions to block, leading to time outs or hanging web pages and frustrated users.

`A` reads `documents`

postgres=*# SELECT document FROM documents;
  document
-------------
 Imagine
 a
 bunch
 of
 interesting
 documents
(6 rows)

At this point, A starts doing a bunch of indexing into a search engine, and it's going to take a long time.

`B` attempts to take an `AccessExclusiveLock`

postgres=# BEGIN;
BEGIN
postgres=*# ALTER TABLE documents ADD COLUMN is_indexed boolean;

This transaction is blocked because A has some locks that are in conflict with the AccessExclusiveLock on documents that B is trying to take. We can check pg_locks to see what the status is right now:

postgres=# SELECT 
  l.relation::regclass AS object_name, 
  l.mode, 
  locktype, 
  granted
FROM pg_locks l
JOIN pg_class c ON c.oid = l.relation
JOIN pg_namespace n ON n.oid = c.relnamespace
AND n.nspname != 'pg_catalog';
  object_name   |        mode         | locktype | granted
----------------+---------------------+----------+---------
 documents_pkey | AccessShareLock     | relation | t
 documents      | AccessShareLock     | relation | t
 documents      | AccessExclusiveLock | relation | f

The two AccessShareLocks are held by A, which is slow and B is waiting for the AccessExclusiveLock.

`C` attempts to read one row from `documents`

postgres=# BEGIN;
BEGIN
postgres=*# SELECT document FROM documents where id = 3;

C is now blocked, waiting to acquire an AccessShareLock:

postgres=# SELECT 
  l.relation::regclass AS object_name, 
  l.mode, 
  locktype, 
  granted
FROM pg_locks l
JOIN pg_class c ON c.oid = l.relation
JOIN pg_namespace n ON n.oid = c.relnamespace
AND n.nspname != 'pg_catalog';
  object_name   |        mode         | locktype | granted
----------------+---------------------+----------+---------
 documents_pkey | AccessShareLock     | relation | t
 documents      | AccessShareLock     | relation | t
 documents      | AccessExclusiveLock | relation | f
 documents      | AccessShareLock     | relation | f
(4 rows)

At this point, B can't proceed because of A and C can't proceed because of B. If B aborts and rolls back, all instances of C will be able to proceed. If A is fast enough, it might be that nobody notices anything, if it's also true that B is fast. If A is slow enough, some people might start getting alerts and start an investigation before A commits.

Mitigations

Scenario 1 would be resolved if we made A from that scenario into 2 migrations; the SELECT document FROM documents WHERE id = 3 transaction does not conflict with the locks that UPDATE documents SET is_indexed = true takes.

For the common example of wanting to add a NOT NULL-column, we can add the column without the NOT NULL constraint, but with the default value in a migration. This is pretty much instant. Then we can issue a second migration to populate the value (it's safer still if we can do this in small batches to avoid long-running transactions), then a third migration to add the NOT NULL constraint.

Scenario 2 needs a different strategy. Maybe we can design the job so we can solve the problem by adding a new documents_index_status table that has a foreign key to documents?

Once you know these patterns, it's easy to test how possible mitigations would work out by practicing with concurrent sessions and see what works and what blocks. What would be very nice however, would be to have some tooling that told you that you were about to do something potentially dangerous.

Detection in CI/CD

Recently I learned from a colleague that flyway has a very cool callback feature that will let you run queries in the same transaction as your migrations, and I'm realizing that there is a lot of potential to build nice tooling with this feature. I have taken a stab at detecting migrations that take AccessExclusiveLocks on tables that aren't new. I do this in three steps.

Snapshot `pg_class` before migration

When flyway gives my callback the BEFORE_EACH_MIGRATE event, I take a snapshot of pg_class to temporarily store which relations that existed before the migration ran:

CREATE TABLE pg_class_before_migrate as SELECT * FROM pg_class;

The reason we do this is that we don't want to yell wolf when migrations create new tables or indexes, since the locks only matter if other transactions try to use the relations concurrently with the migration.

Find relation level `AccessExclusiveLock`s

On the AFTER_EACH_MIGRATE event, I retrieve all relation level locks that my transaction has taken on database objects that existed before the migration ran. If the migration has issued a SET ROLE we may need to change role back to the one that created the pg_class_before_migrate-table.

SELECT 
  n.nspname::text AS schema_name, 
  l.relation::regclass::text AS object_name
FROM pg_locks l
  JOIN public.pg_class_before_migrate c ON c.oid = l.relation
  JOIN pg_namespace n ON n.oid = c.relnamespace
WHERE l.locktype = 'relation'
  AND l.mode = 'AccessExclusiveLock'
  AND l.pid = pg_backend_pid(); -- Locks held by current transaction

Clean up

We probably shouldn't leave that snapshot of pg_class hanging around forever:

DROP TABLE pg_class_before_migrate;

At this point, you can choose to have the CI/CD comment on the pull request that the migration will lock documents and therefore you should check what types of transactions that may try to use that table concurrently with the migration. Possibly have some LLM generate a helpful text about what could go wrong with the migration script? Maybe match the locks against a list of tables that are acceptable to lock, or a list of tables that are known to be dangerous to lock?

Future work

Note that there are lots of corner cases still here, I'm very enthusiastic about lots of ideas in this space and definitely want to write and share some more code once I know more about what's possible.

Taking a snapshot of pg_class potentially also gives us another few things we can detect. If we can run these migrations against something like an anonymized copy of the production database, we may be able to issue ANALYZE before and after the migration to detect some table rewrites -- such an operation should change the relationship between reltuples and relpages in pg_class, by increasing the size of the tuples. That might also allow us to flag migrations that rewrite large tables. We can also snapshot other tables from information_schema, or tables like pg_attribute to build a pretty good data structure describing the effect of the migration in terms of both schema and content.

Errata

It is no longer true since postgres 11 that adding a not null column with a default value causes a table rewrite, unless the default value is VOLATILE, like generating a random uuid or fetching a timestamp.

This post was originally posted on kaveland.no.

Scenario 1

A adds a column to documents

A starts a slow statement

B attempts to read a single document