[Apache][Doris][Feature][FE]Support users with high QPS in the same query through result set cache

[songenjie]

-Support users with high QPS in the same query through result set cache

Fixes #4283

[yangzhg]

you should also add a trigger that when a table has load or delete ops, you should refresh all cache related that table

[kangkaisen]

@songenjie Hi, what's your design doc for this PR? I think this PR is repeated with SQL mode cache in #4330 , what do you think it？

[songenjie]

@songenjie Hi, what's your design doc for this PR? I think this PR is repeated with SQL mode cache in #4330 , what do you think it？

@kangkaisen

1. Sql Cache Mode

• If this switch is turned on, the SQL query result set will be cached. If the interval between the last visit version time in all partitions of all tables in the query is greater than cache_last_version_interval_second, and the result set is less than cache_result_max_row_count, the result set will be cached, and the next same SQL will hit the cache.

2. this PR Result Cache

• Solve the same connect ID, the same SQL high QPS problem, the same user, in result_ cache_ expire_ after_ In_ Within milliseconds time, executing the same SQL will hit the result cache

[kangkaisen]

1. this PR Result Cache

• Solve the same connect ID, the same SQL high QPS problem, the same user, in result_ cache_ expire_ after_ In_ Within milliseconds time, executing the same SQL will hit the result cache

Hi, I think the selfsame OLAP querys in prod env is very rare. Do you test this PR in your prod env? what's the cache hit rate?

The Apache Kylin use SQL result cache Like this PR, After Kylin run many years in MEITUAN, we could confirm the SQL result cache is meaningless for OLAP system.

The Apache Druid use segment cache like #4330, After Druid run many years in MEITUAN, we could confirm the segment cache is useful for OLAP system.

So I think If we already have segment query cache, we don't need the query result cache.

[songenjie]

Hi, I think the selfsame OLAP querys in prod env is very rare. Do you test this PR in your prod env? what's the cache hit rate? The Apache Kylin use SQL result cache Like this PR, After Kylin run many years in MEITUAN, we could confirm the SQL result cache is meaningless for OLAP system.

3ks

the result cache has been used in Jingdong for more than half a year, and the cache hit rate is also very high, especially in the e-commerce industry like ours, in special periods and business scenarios, its hit rate is very high

As far as I know, most of the business of meituan is near real-time, and the scope of business use and analysis is slightly different from ours. So maybe the hit rate of cache is not high

[morningman]

I didn't see any code about cache invalid? Did I miss something?

[songenjie]

I didn't see any code about cache invalid? Did I miss something?

https://github.com/apache/incubator-doris/blob/6c2ca4690139acdf917193ec4d0acda39b7a148b/fe/fe-core/src/main/java/org/apache/doris/qe/StmtExecutor.java#L635

[morningman]

I didn't see any code about cache invalid? Did I miss something?

https://github.com/apache/incubator-doris/blob/6c2ca4690139acdf917193ec4d0acda39b7a148b/fe/fe-core/src/main/java/org/apache/doris/qe/StmtExecutor.java#L635

I mean, where is the logic for judging cache invalidation? The cache key (NamedKey) is only composed by connection id and origin stmt? What if the data has been unchanged?

[songenjie]

I mean, where is the logic for judging cache invalidation? The cache key (NamedKey) is only composed by connection id and origin stmt? What if the data has been unchanged?

https://github.com/apache/incubator-doris/blob/dd42c6b580afcbb53436dd8b531e03e5fa5d01ed/fe/fe-core/src/main/java/org/apache/doris/cache/SimpleLocalCache.java#L63

use Caffeine.newBuilder().recordStats();

[morningman]

I mean, where is the logic for judging cache invalidation? The cache key (NamedKey) is only composed by connection id and origin stmt? What if the data has been unchanged?

https://github.com/apache/incubator-doris/blob/dd42c6b580afcbb53436dd8b531e03e5fa5d01ed/fe/fe-core/src/main/java/org/apache/doris/cache/SimpleLocalCache.java#L63

use Caffeine.newBuilder().recordStats();

So it only depends on time? If data changed within Config.result_cache_expire_after_in_milliseconds, user in same context still read the old data?

[songenjie]

I mean, where is the logic for judging cache invalidation? The cache key (NamedKey) is only composed by connection id and origin stmt? What if the data has been unchanged?

https://github.com/apache/incubator-doris/blob/dd42c6b580afcbb53436dd8b531e03e5fa5d01ed/fe/fe-core/src/main/java/org/apache/doris/cache/SimpleLocalCache.java#L63

use Caffeine.newBuilder().recordStats();

So it only depends on time? If data changed within Config.result_cache_expire_after_in_milliseconds, user in same context still read the old data?

right