[Question] TTL expressions allow non-deterministic functions

[dcelasun]

Company or project name

CloudQuery

Describe what's wrong

This is supposed to be disallowed since ClickHouse/ClickHouse#51858, but I noticed that now() is silently allowed. So I tried reproducing it on fiddle using rand():

CREATE TABLE users (uid Int16, ts DateTime) ENGINE=MergeTree ORDER BY uid
TTL ts + INTERVAL 1 SECOND DELETE WHERE ts>0 and rand()<2147483647;
insert into users(uid, ts) values (123, now()-interval 5 second);
select * from users;
optimize table users final;
select * from users;

Run that a few times and observe that the row isn't deleted ~50% of the time. I've checked that allow_suspicious_ttl_expressions isn't enabled.

Does it reproduce on the most recent release?

Yes

How to reproduce

• Run the above fiddle.
• Observe that the row isn't deleted ~50% of the time.

Expected behavior

Either rand() (and now(), now64() etc.) shouldn't be allowed, or they should be explicitly supported.

Error message and/or stacktrace

No response

Additional context

No response

[alexey-milovidov]

We don't allow expressions that don't depend on any columns of the table, but still allow non-deterministic expressions. If you want this to be implemented, you can create a feature request.

[dcelasun]

We don't allow expressions that don't depend on any columns of the table, but still allow non-deterministic expressions. If you want this to be implemented, you can create a feature request.

I'm not asking for it to implemented. My example depends on a column of the table (ts) and uses a non-deterministic function and Clickhouse allows it. I'm saying it shouldn't be allowed.

[alexey-milovidov]

Yes, probably. However, everything will work as specified even with a non-deterministic TTL.

If merges are processed concurrently on replicas, one replica will complain about a different checksum after the merge and resort to downloading the data part from another replica.

[dcelasun]

So, you don't consider this a bug? Even though the linked PR clearly tries (and apparently fails) to disallow non-deterministic functions in TTLs?

I feel like, at a minimum, this is a documentation bug. Not a question that was asked and answered, like your labels suggest.