public inbox for [email protected]  

help / color / mirror / Atom feed

From: Benjamin Coutu <[email protected]>
To: PostgreSQL Hackers <[email protected]>
Cc: Tom Lane <[email protected]>
Cc: Robert Haas <[email protected]>
Cc: David Rowley <[email protected]>
Cc: Peter Geoghegan <[email protected]>
Subject: Re: disfavoring unparameterized nested loops
Date: Thu, 29 Sep 2022 16:32:58 +0200
Message-ID: <[email protected]> (raw)

I'd like to revamp this important discussion.

As is well described in this fairly recent paper here https://www.vldb.org/pvldb/vol9/p204-leis.pdf (which also looks at Postgres) "estimation errors quickly grow as the number of joins increases, and that these errors are usually the reason for bad plans" - I think we can all get behind that statement.

While nested loop joins work great when cardinality estimates are correct, they are notoriously bad when the optimizer underestimates and they degrade very fast in such cases - the good vs. bad here is very asymmetric. On the other hand hash joins degrade much more gracefully - they are considered very robust against underestimation. The above mentioned paper illustrates that all mayor DBMS (including Postgres) tend to underestimate and usually that underestimation increases drastically with the number of joins (see Figures 3+4 of the paper).

Now, a simple approach to guarding against bad plans that arise from underestimation could be to use what I would call a nested-loop-conviction-multiplier based on the current depth of the join tree, e.g. for a base table that multiplier would obviously be 1, but would then grow (e.g.) quadratically. That conviction-multiplier would *NOT* be used to skew the cardinality estimates themselves, but rather be applied to the overall nested loop join cost at each particular stage of the plan when comparing it to other more robust join strategies like hash or sort-merge joins. That way when we can be sure to have a good estimate at the bottom of the join tree we treat all things equal, but favor nested loops less and less as we move up the join tree for the sake of robustness.
Also, we can expand the multiplier whenever we fall back to using the default cardinality constant as surely all bets are off at that point - we should definitely treat nested loop joins as out of favor in this instance and that could easily be incorporated by simply increasing the conviction-mutliplier.

What are your thoughts on this simple idea - is it perhaps too simple?

Cheers, Ben

view thread (59+ messages)  latest in thread

reply

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Reply to all the recipients using the --to and --cc options:
  reply via email

  To: [email protected]
  Cc: [email protected], [email protected], [email protected], [email protected], [email protected], [email protected]
  Subject: Re: disfavoring unparameterized nested loops
  In-Reply-To: <[email protected]>

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

This inbox is served by agora; see mirroring instructions
for how to clone and mirror all data and code used for this inbox