Query wrong result 18.14.8(9) (CH cashes ? subquery result)

[den-crane]

Preparing:

CREATE TABLE dt(tkey Int32) ENGINE = MergeTree order by tuple();
insert into dt values (300000);
CREATE TABLE testx(t Int32, a UInt8) ENGINE = MergeTree ORDER BY tuple();
INSERT INTO testx VALUES (100000, 0);
select count(*) from testx where not a and t < (select tkey from dt);
1
drop table dt;
CREATE TABLE dt(tkey Int32) ENGINE = MergeTree order by tuple();
insert into dt values (0);

BUG:
select count(*) from dw.testx where not a and t < (select tkey from dt);
1
select tkey from dt
0
select count(*) from dw.testx where not a and t < 0
0

So CH does not see that value in the table dt has changed from 300000 to 0.

optimize_move_to_prewhere 0
enable_optimize_predicate_expression 0

[den-crane]

Workaround: move all other conditions to prewhere and leave the condition with subquery in where section.

select count(*) from testx
prewhere not a
where t < (select tkey from dt);

[den-crane]

select count(*) from dw.testx where not a and t < (select tkey from dt);
┌─count()─┐
│ 1 │
└─────────┘
1 rows in set. Elapsed: 0.002 sec.

set compile_expressions = 0;

select count(*) from dw.testx where not a and t < (select tkey from dt);
0 rows in set. Elapsed: 0.002 sec.

[den-crane]

In our system we don't drop/create dt.
It's a replacingmergetree table and it's data is different for every query.
And more complicated
time_key < (select time_key from dim_time where time_stamp = (select max(last_record_time) from data_status_current))
dim_time is constant, data_status_current has different max data every time.

[alesapin]

Den, I can't reproduce the problem from your example:
My steps:

:) CREATE TABLE dt(tkey Int32) ENGINE = MergeTree order by tuple();

Ok.

0 rows in set. Elapsed: 0.002 sec. 
:)  insert into dt values (300000);

Ok.

1 rows in set. Elapsed: 0.002 sec.
:)  CREATE TABLE testx(t Int32, a UInt8) ENGINE = MergeTree ORDER BY tuple();

Ok.

0 rows in set. Elapsed: 0.002 sec. 
:)  INSERT INTO testx VALUES (100000, 0);

Ok.

1 rows in set. Elapsed: 0.002 sec.
:)  select count(*) from testx where not a and t < (select tkey from dt);

┌─count()─┐
│       1 │
└─────────┘

1 rows in set. Elapsed: 0.004 sec. 

:) drop table dt;

Ok.

0 rows in set. Elapsed: 0.001 sec. 

:)  CREATE TABLE dt(tkey Int32) ENGINE = MergeTree order by tuple();

Ok.

0 rows in set. Elapsed: 0.003 sec.

:)  insert into dt values (0);

Ok.

1 rows in set. Elapsed: 0.002 sec. 
:)  select count(*) from testx where not a and t < (select tkey from dt);

┌─count()─┐
│       0 │
└─────────┘

1 rows in set. Elapsed: 0.004 sec. 

In last select query you use dw.testx table instead of testx, maybe this is the cause of wrong answer?

About expressions compilation:

1. Expressions are compiled when 3 identical expressions occur in queries. If you queried your select less then three times, then it's not compile expressions problem.
2. There is special asynchronous_metric in table system.asynchronous_metrics called CompiledExpressionCacheCount. Even after three times I called this expression metric doesn't increment. It means that this expression is too simple to be compiled.

Maybe you can provide more complex example?

[den-crane]

select name, value from system.settings where changed;

┌─name──────────────────────────────────────┬─value────────────┐
│ connect_timeout_with_failover_ms │ 1000 │
│ use_uncompressed_cache │ 1 │
│ optimize_move_to_prewhere │ 0 │
│ load_balancing │ nearest_hostname │
│ compile_expressions │ 0 │
│ distributed_aggregation_memory_efficient │ 1 │
│ skip_unavailable_shards │ 1 │
│ http_connection_timeout │ 15 │
│ empty_result_for_aggregation_by_empty_set │ 1 │
│ max_bytes_before_external_group_by │ 25374640128 │
│ max_memory_usage │ 50749280256 │
└───────────────────────────────────────────┴──────────────────┘

[den-crane]

it's tricky and works with engine = memory and sometimes caches other value

[alesapin]

Can you provide select * from system.asynchronous_metrics ?

[den-crane]

[den-crane]

SELECT *
FROM system.asynchronous_metrics

┌─metric──────────────────────────────────┬──────value─┐
│ jemalloc.background_thread.run_interval │          0 │
│ jemalloc.background_thread.num_runs     │          0 │
│ jemalloc.background_thread.num_threads  │          0 │
│ jemalloc.retained                       │  557051904 │
│ jemalloc.mapped                         │ 2941521920 │
│ jemalloc.resident                       │ 2835759104 │
│ jemalloc.metadata_thp                   │          0 │
│ jemalloc.metadata                       │   15238536 │
│ jemalloc.allocated                      │ 2706048264 │
│ UncompressedCacheCells                  │        110 │
│ MarkCacheFiles                          │        110 │
│ ReplicasSumInsertsInQueue               │          0 │
│ MarkCacheBytes                          │       1760 │
│ UncompressedCacheBytes                  │    1420133 │
│ CompiledExpressionCacheCount            │          4 │
│ Uptime                                  │        723 │
│ ReplicasMaxQueueSize                    │          0 │
│ ReplicasMaxInsertsInQueue               │          0 │
│ ReplicasMaxMergesInQueue                │          0 │
│ ReplicasMaxRelativeDelay                │          0 │
│ jemalloc.active                         │ 2719969280 │
│ CompiledExpressionCacheBytes            │       4560 │
│ ReplicasSumQueueSize                    │          0 │
│ ReplicasMaxAbsoluteDelay                │          0 │
│ ReplicasSumMergesInQueue                │          0 │
│ MaxPartCountForPartition                │         10 │
└─────────────────────────────────────────┴────────────┘

[alesapin]

yes, reproduced with:

optimize_move_to_prewhere 0
enable_optimize_predicate_expression 0

[den-crane]

yes, I think optimize_move_to_prewhere 1
transform it to
select count(*) from testx prewhere t <(subquery) where not a

[alesapin]

Yes. This problem seems to be very special case. Сondition t < (subquery) executes only when subquery returns single element.

[alesapin]

https://github.com/yandex/ClickHouse/pull/3430/files hotfix

[den-crane]

t.me/clickhouse_ru/71285

ch0 :) select eventDate, count() from Events where eventDate = today()-1 group by eventDate settings compile_expressions = 0;

┌──eventDate─┬──count()─┐
│ 2018-10-22 │ 70056042 │
└────────────┴──────────┘

1 rows in set. Elapsed: 0.069 sec. Processed 70.06 million rows, 140.12 MB (1.02 billion rows/s., 2.04 GB/s.)

ch0 :) select eventDate, count() from Events where eventDate = today()-1 group by eventDate settings compile_expressions = 1;

┌──eventDate─┬─count()─┐
│ 2018-10-21 │ 758 │
└────────────┴─────────┘

1 rows in set. Elapsed: 0.017 sec. Processed 70.06 million rows, 140.12 MB (4.13 billion rows/s., 8.27 GB/s.)

[alesapin]

We disable compile expressions by default in 14.10.

[alesapin]

Seems to be fixed in master #3457. But we will test this feature more carefully.