UDDSketch assertion at crates/udd-sketch/src/lib.rs:290:9

[tyhoff]

Relevant system information:

• OS: Aiven's TimescaleDB provider
• PostgreSQL version, 13.6
• TimescaleDB Toolkit version: 1.5.1
• Installation method: Aiven

Describe the bug
I hit an assert when I don't expect to when creating a UDDSketch.

To Reproduce
Apologies for the seemingly random test case, but here it goes.

CREATE TABLE "public"."timescale_bug" ("n" float8 NOT NULL);

Load the attached CSV file into the table.

Run

SELECT
	uddsketch(100, 0.01, CAST(n AS FLOAT))
FROM
	timescale_bug

Expected behavior
Obvious

Actual behavior

Query 1 ERROR: ERROR:  assertion failed: (self.gamma.powf(1.0 / f64::powi(2.0, self.compactions as i32)) -\n     other.gamma.powf(1.0 / f64::powi(2.0, other.compactions as i32))).abs() <\n    f64::EPSILON
CONTEXT:  crates/udd-sketch/src/lib.rs:290:9

Fudging with the bucket count or error will get it to work, but it's random when the assert is tripped.

timescale_bug.csv.zip

[rtwalker]

Thanks for the report! Confirming I can reproduce (on about 1 out of every 10 queries) on 1.5.1 and also when building and installing the main branch.

Will let you know what I find!

[rtwalker]

explain select uddsketch(100, 0.01, n::float)::text from timescale_bug;
                                           QUERY PLAN                                           
------------------------------------------------------------------------------------------------
 Finalize Aggregate  (cost=9233.67..9233.68 rows=1 width=32)
   ->  Gather  (cost=9233.44..9233.65 rows=2 width=32)
         Workers Planned: 2
         ->  Partial Aggregate  (cost=8233.44..8233.45 rows=1 width=32)
               ->  Parallel Seq Scan on timescale_bug  (cost=0.00..7343.15 rows=356115 width=8)

[bvanelli]

What is the status on this one? Would an upgrade to > 1.60 solve this issue? I'm having the same bug.

[rtwalker]

We haven't included a fix for this bug yet. From what I can tell, it appears likely that at least one of the assertions we are making when combining the UDDSketches from different parallel workers is too sensitive to floating point arithmetic.

I will note that in testing this out, I haven't been able to trigger this bug after running
SET max_parallel_workers_per_gather = 0;

I'll be sure to update this issue once we've released a bug fix.

[tyhoff]

@rtwalker Thanks for digging into this one! Excited to try it out in the next release.

[jerryxwu]

@tyhoff If you haven't noticed, we've just included the fix in the latest 1.8.0 release. We'd love to hear whether it's fixed for you or not.

[tyhoff]

@tyhoff If you haven't noticed, we've just included the fix in the latest 1.8.0 release. We'd love to hear whether it's fixed for you or not.

@jerryxwu I noticed! I'm still waiting for it to be published to Aiven (Timescale Managed) sadly.