When we have a vector of 2048 dimension，the get_by_id is slow

[xealml]

I want to get a vector by id in milvus, but the request time is too long.
the log as below.
PS: 600 * 10000 entities.

[2020-12-11 14:37:21,227][DEBUG][SERVER][PrintTimeRecord][reqsched_thread] SegmentReader::LoadEntities: multi_modal_embedding: totally cost (4.199081 seconds [4199.081138 ms])
[2020-12-11 14:37:21,233][DEBUG][SERVER][PrintTimeRecord][reqsched_thread] SegmentReader::LoadEntities: p_type: totally cost (0.005937 second [5.937437 ms])
[2020-12-11 14:37:21,239][DEBUG][SERVER][PrintTimeRecord][reqsched_thread] SegmentReader::LoadEntities: p_id: totally cost (0.005831 second [5.831457 ms])
[2020-12-11 14:40:17,256][DEBUG][SERVER][PrintTimeRecord][reqsched_thread] SegmentReader::LoadEntities: multi_modal_embedding: totally cost (4.244736 seconds [4244.735509 ms])
[2020-12-11 14:40:17,260][DEBUG][SERVER][PrintTimeRecord][reqsched_thread] SegmentReader::LoadEntities: p_type: totally cost (0.004836 second [4.835718 ms])

[shiyu22]

Hi,
Which version are you using? Actually, when you use get_enity_by_id, Milvus will read data from disk, so the process is not fast

[yhmo]

This is a performance issue of 0.11.0
In 0.10.x, to get a vector from collection, milvus only read few bytes by using seek() and read(), it only cost several milliseconds to get a vector
In 0.11.0, get_entity_by_id() load whole segment from disk into memory. Each segment row count default is 512*1024, for 2048 dimensional vectors, each segment size is 4GB, that means every time get_enity_by_id() is called, it will read a 4GB file into memory.

[yhmo]

I have create issue #4488 for this issue, I suppose fix it in 0.11.1