mirror of
https://github.com/ClickHouse/ClickHouse.git
synced 2024-10-19 15:01:03 +00:00
eae56e2d02
If within one INSERT the background worker will flush the buffer, then it will steal the allocation, and if it will steal enough then max_memory_usage for the INSERT query context could be exceeded. Fix this by forcing the buffer flush from the query context.
38 lines
1.3 KiB
SQL
38 lines
1.3 KiB
SQL
DROP TABLE IF EXISTS null_;
|
|
DROP TABLE IF EXISTS buffer_;
|
|
|
|
-- Each UInt64 is 8 bytes
|
|
-- So 10e6 rows is 80e6 bytes
|
|
--
|
|
-- Use LIMIT max_rows+1 to force flush from the query context, and to avoid
|
|
-- flushing from the background thread, since in this case it can steal memory
|
|
-- the max_memory_usage may be exceeded during squashing other blocks.
|
|
|
|
|
|
CREATE TABLE null_ (key UInt64) Engine=Null();
|
|
CREATE TABLE buffer_ (key UInt64) Engine=Buffer(currentDatabase(), null_,
|
|
1, /* num_layers */
|
|
10e6, /* min_time, placeholder */
|
|
10e6, /* max_time, placeholder */
|
|
0, /* min_rows */
|
|
10e6, /* max_rows */
|
|
0, /* min_bytes */
|
|
80e6 /* max_bytes */
|
|
);
|
|
|
|
SET max_memory_usage=10e6;
|
|
SET max_block_size=100e3;
|
|
|
|
-- Check that max_memory_usage is ignored only on flush and not on squash
|
|
SET min_insert_block_size_bytes=9e6;
|
|
SET min_insert_block_size_rows=0;
|
|
INSERT INTO buffer_ SELECT toUInt64(number) FROM system.numbers LIMIT toUInt64(10e6+1); -- { serverError 241 }
|
|
|
|
OPTIMIZE TABLE buffer_; -- flush just in case
|
|
|
|
SET min_insert_block_size_bytes=0;
|
|
SET min_insert_block_size_rows=100e3;
|
|
INSERT INTO buffer_ SELECT toUInt64(number) FROM system.numbers LIMIT toUInt64(10e6+1);
|
|
-- Check that 10e6 rows had been flushed from the query, not from the background worker.
|
|
SELECT count() FROM buffer_;
|