The only scalable delete in Postgres is DROP TABLE
planetscale.com
planetscale.com
> Counterintuitively, large DELETEs add work to the database.
There is nothing counterintuitive about this. It takes just as much work to delete a row as it takes to insert a row. Why wouldn't it? Obviously you have to do almost all the same operations: write a log, write the deletion, update indices, replicate it, etc.
And yes, it's a well-known trick for all major relational databases (not just Postgres) that if you want to delete 90% of rows from a large table, it's much faster to just copy the rows you want to keep to a new table, run DROP TABLE on the old table, and rename the new table to the old table. Since DROP TABLE is ~instantaneous, mainly involving table-level metadata.
DELETE scales just fine, in the sense that if you are constantly inserting and deleting individual rows, DELETE scales the same as INSERT.
Basic database functionality is designed around the assumption of lots of small transactions. Whenever you have to do something involving millions of rows at once, you generally need to investigate solutions that work well in "bulk". E.g. loading rows directly from a file rather than with SQL, adding indices only after the data has been loaded rather than before, disabling foreign key checks on large operations (if you know by design that the keys are valid)... and yes, taking advantage of DROP TABLE instead of DELETE. This doesn't mean small transactions aren't scalable, it just means bulk operations are qualitatively different and benefit from their own solutions. And DELETE is no different from INSERT in this regard.
Dumb question but why does the optimizer not just do that in secret then? Seems like something that should be detectable with some heuristics.
You can only do the DROP TABLE trick if you know nothing else is writing to the table at the same time. You know if that's the case, according to your business logic. The database has no idea.
The DROP TABLE trick effectively bypasses all the normal guarantees of data consistency. This is why it's so fast. But you have to know that that's a safe thing to do for your data.
The tricky part is that the latency characteristics of these operations would be pretty surprising and unintuitive. It has the same problems as virtual memory and mark/sweep GC; sometimes, depending on system state and things that other threads are doing, an unrelated operation might block for very long time periods and give you huge user-visible pauses. It's often better to force these expensive operations to be explicit so that the developer has to think through the latency & consistency implications and make the tradeoffs they want.
It takes far more work to delete/update than insert. My recent example is updating ~2TB of text data was about 40x slower than inserting 12TB (was trying to correct some large text truncation that occurred during migration into PG, ended up being faster to redo).
Updating rows of text data is going to be more work, because variable-length text can't be updated in-place. So in terms of allocating space, it's more like a delete plus an insert. That's not surprising. (An in-place update that doesn't touch indices is generally going to be faster than an insert, though.)
I'm not aware of instances where a delete is "far more work" than an equivalent insert though. That's not the general case, and I'm having a hard time thinking of any situations where that would be true.
Unless you're using zHeap, you have a narrow Heap-only-Tuples scenario where the indexes stay the same. TOAST kinda helps there, if the update is off the tuple area itself. The original zHeap docs have a lot of detail about why an UNDO log can help with long running transactions from the past etc.
That is a postgresql specific thing though. Mysql indexes were created with the idea of different storage engines in mind, so Mysql doesn't suffer from the index update ovehead on update/delete the same way.
Uber had a long blog post about switching to Mysql from Postgres for wide tables with hundreds of indexes. The HN entry is still there[1], but I can't read the original post now.
As a side note, I've used postgres partitions to the same effect to drop old data periodically - detach and then drop the partition instead of a direct DELETE (similar tricks in HBase existed).
The post on IA - https://web.archive.org/web/20160304013342/https://eng.uber....
Transactionally across related items with constraints it can explode fast.
If you've ever used FoundationDB this rapidly becomes the defining PITA due to the transaction size limits. Adding/inserting/updates are all far more predictably bounded.
This is precisely where you're going wrong. The insert is upper boundable in advance (you know the set of everything you might potentially have to insert), the delete isn't because you don't know what's in the db until you look.
I strongly recommend poking around with Foundation for this, because it becomes clear that this problem is the defining flaw with the way they tried to architect with layers, to the point they have a queuing system for processing large jobs of this type.
A concurrent DML happening then suddenly your MERGE INTO WHEN NOT MATCHED INSERT/INSERT INTO SELECT is way larger that you thought? I thought "some workloads can suddenly be way larger that I expected" was supposed to be a thing in all non-trivial DML.
If we're still talking postgres, it doesn't update in-place. Update is implemented as delete+insert (where delete is updating metadata so the row is still around for still-running transactions but invisible to future transactions).
> disabling foreign key checks on large operations
And you have to know that, according to your business logic, what you're doing is safe.
Technically correct, but for a small table with a high churn rate, the performance characteristics may be surprising in that the "n" in most big-O calculations includes all inserts since the last VACUUM, not the actual number of resident rows.
Because e.g. DROP also effectively deletes the rows but takes way way less work.
Because your data structure/algorithm supports fast deletes? File systems support deleting entire directories instantly. I'm not aware of any fundamental reason why DELETE in a SQL database must take as long as an insert?
And file systems deleting huge directories instantly is the equivalent of DROP TABLE here, which I also mention in my comment.
What file system supports this? I’ve never experienced the joy of deleting an entire large directory and seeing it disappear instantly.
I’m not sure how this would work with any modern file system that supports basic notions like hard links.
I sincerely hope that Planetscale’s efforts succeed long-term to shift devs’ understanding and acceptance of RDBMS operations. Their blog posts and docs are generally quite good. IME, devs (and even ops-ish teams) simply do not care about all of this, and will create elaborate bespoke tooling to run DELETEs in bulk, because they either don’t understand the capabilities of the database, or don’t want to deal with the [minor] increased complexity that a partitioned schema brings, and will happily pay the extra cost / latency for deletions.
DELETE with well-tuned autovacuum works pretty well. Have seen it work at TBs scale with no hicuups. If DELETEs are large, we used to recommend customers to follow that with a manual VACUUM for table to reclaim space right away for future rows.
DROP TABLE can be risky, it requires an ACCESS EXCLUSIVE LOCK and if its waiting, it blocks all other statements following it, because of how lock queues work in Postgres. And you cannot keep doing high concurrent DROP TABLEs to run your large scale CRUD app.
> And you cannot keep doing high concurrent DROP TABLEs to run your large scale CRUD app
The entire premise here is really about time-series workloads where most operations are based on a timestamp. In those apps partition dropping has been a standard and recommended retention strategy for years. That's precisely why extensions like pg_partman and TimescaleDB exist. Given that context, the title feels more clickbaity, and could easily mislead readers into thinking this applies broadly to OLTP systems when it doesn't;
UUID v7 to the rescue!
Similarly, you often have to remind devs that in many databases an UPDATE is just an INSERT + DELETE, with all of the scaling issues implied.
Don't get me wrong, I've definitely done it before, but it's in the same bucket as VACUUM for me... high impact interventions used to fix a mistake I made, not "course of business" actions.
If you mean your database is seeing extended periods of no updates to a table, you still want to vacuum, maybe even vacuum full if you know when traffic stopped and for how long to get the best possible read performance.
If you have quiet periods in both reads and writes, enjoy the luxury of having an unused database to operate.
We use materialized views for append-only timeseries data for https://lobu.ai and the retention policies define how we DROP the tables so we don't DELETE/UPDATE any rows in the tables.
The long term storage is Iceberg on S3 that's ingested via Postgresql replication, suitable for OLAP use-cases. Postgresql only stores the dimensional OLTP data the users can update and the hot append-only event data.
https://github.com/pgpartman/pg_partman/blob/development/doc...
The better approach is either to change your storage engine (e.g. OrioleDB is working on adding the undo log to Pg), or to shard which distributes the vacuum load across multiple servers.
been exploring clickhouse and while it is definitely not a general purpose DB, for time-series shaped data that can survive some insert latency, the automatic partition-based TTL is very nice and, at least so far, requires zero attention to maintain
which I guess is solved by `pg_partman` at the bottom of the post
I have used a very similar strategy by forking repack client https://github.com/reorg/pg_repack/pull/326 This works out of the box with rds/cloudsql etc.
Deletes are Writes and Writes are resource intensive. This is more prominent on databases like Elastic Search.
When I was tasked to delete millions of (old) documents, it overloaded the cluster and almost brought it down. Only scalable solution was to split the index and drop the whole index.
I am not sure of this is the best way to do this, but it works for me.
One day you'll maybe discover ACID properties of RDBMS systems, and all the puzzles will fall into place.
- We wrote a cronjob to periodically DELETE for a retention policy on a table we'd just created. Most senior person on the team reviewed it, looked fine.
- Unusually for us, we prioritize QA'ing a different feature for release, delaying the release of this cronjob and a bunch of other code.
- During that delay, the new table accumulated many times more rows to be deleted than we'd expected during review.
- Release happens. All looks well since the initial delete wasn't a migration and cronjob hasn't run yet; engineer doing the release signs off.
- Cronjob runs, deleting hundreds of millions of rows quickly.
- Next day, replica lag's high and MySQL's transaction history is very high. MySQL keeps transaction history around until purge threads have visited all the affected pages on disk.
- The bad cluster conditions last for days and lead to other problems.
This omits detail and the 'noise' of everything else we were watching. But it gets across how the code and MySQL behaved.
Like most exciting events, it led to multiple changes to avoid a repeat. For retention policies, our new approach was one at the end of PlanetScale's post, to partition and drop old partitions. Transitioning to this from a huge unpartitioned table can be fun!
If a table is append-only and already huge, with lots of rows already past the retention threshold, you might only copy the rows to be kept to the new partitioned table: copy what you can, lock tables, do a last catch-up copy and swap tables. (Roughly the blog's 'performant one-off delete'.)
If the table's merely kind of big, gh-ost or such could allow you to ALTER without causing lag, locking, etc.
At a scale below that, you could run a slow incremental 'nibble' delete while watching server stats, and a step below that, plain ALTERs or DELETEs are fine.
Using partitioning has fun bits, too. In MySQL, the partition key has to be part of any unique index, understandably. But you have to keep that in mind when you're using INSERT..ON DUPLICATE KEY UPDATE and relying on uniqueness to trigger the update. Things stay interesting!
I hear Vitess shops like PlanetScale usually don't run multi-terabyte myqsld instances in the first place: even when physical nodes are big, they run many smaller mysqlds on them. That wouldn't make all this fully irrelevant--huge deletes would still sometimes be worse than copy-swap-drop--but it does seem real handy for taming issues that tend to worsen with mysqld size, like replication lag. All to say, little bit jelly of their setup over there!
The main reason was to avoid a separate cron job, but it had other benefits (and downsides) too. Something like:
DELETE FROM foo WHERE expires_at < now() LIMIT 10;
INSERT INTO foo .....;
And by deleting (up to) 10 each time we insert one, it ensures obsolete things will eventually get deleted.
Obviously, this isn't viable when the deletions must happen due to strict policy (e.g. legal compliance) since it can't ensure when things get deleted, just that they eventually do. IIRC, in my case, I used it for a password reset tokens table. There's no legal issue there and keeping expired ones around is fine as long as the code also checks `expires_at` to make sure it's still valid (which would be a good practice regardless, for defense in depth).
I have no idea why, it seems such an obvious feature for supporting large databases.
Something like:
WITH ids AS (
SELECT id FROM foo
WHERE ...
LIMIT 10
)
DELETE FROM foo
USING ids
WHERE foo.id=ids.id;
DELETE FROM foo
WHERE id IN (SELECT id FROM foo WHERE ... LIMIT 10);Storing some data in a binary file isn't very hard. Making it so that you can do quick lookups on it (indexes) and implementing joins in a sane way is kinda hard, but easy compared to the real problem:
Ensuring ACID (in the case of "traditional" databases). I.e. Atomicity, Consistency, Isolation, Durability.
You need to protect against data corruption in the event of failure, all while guaranteeing atomic operations at the user level concurrently (in most production DBs; SQLite is a notable exception in that it fully serializes writes --- but it can get away with this because it's an embedded database with the primary use case of a single-process writer). And the entire thing must land on a known good state at the end of all of those concurrent transactions.
... and they must do it all while maintaining good performance, and sometimes on a combination of filesystem + hardware that's actively hostile towards the idea of data integrity (e.g. hidden RAM caches in disk or RAID controllers that don't flush on power loss --- thankfully, those are getting rarer, or so I've come to understand).