Grid Store pre reserving any space??

[narendra9079]

Is any space is reserved for the grid store when collection is created ??

[timvisee]

Yes.

Many more components use/reserve some disk space, including RocksDB.

[narendra9079]

Thank you, just one more thing, for rocksdb, the rocksdb flag is there so by default which one is used when i insert the points gridstore or rocksdb?? if i am using version 1.16.0 code.

Also what is the meaning of below data is this much space is reserved for payload storage in each segment when collection is created ??

{"page_size_bytes":33554432,"block_size_bytes":128,"region_size_blocks":8192,"compression":"LZ4"}

[timvisee]

In 1.16.0 Gridstore is used by default.

It is possible that you still have some RocksDB components if you upgraded to 1.16.0 from an older version. That is only possible if you have not been actively updating your collections. Qdrant 1.16.1 force migrates everything away from RocksDB on startup.

[narendra9079]

Thank you, could you please tell something about the data which i shared ??

[narendra9079]

one more follow up the default feature in segment cargo.tomal in version 16.0 is rocksdb then how are you using gridstore by deafult ??

[timvisee]

The page size is relevant, so 33554432 bytes are reserved on disk for Gridstore in a segment.

one more follow up the default feature in segment cargo.tomal in version 16.0 is rocksdb then how are you using gridstore by deafult ??

It is still compiled into the binary so that Qdrant can open old storage formats. It is not actively used.

[narendra9079]

okay thank you.

[narendra9079]

Could you please tell why did u set default size of gridstore page 32 MB?? If I am more concerned about memory should i use rocksdb or gridstore, because if i set gridstore page size low then it will increase my latency by creating new pages every time, and if i set a high value it will unnecessary occupy my disk space, If i dont know in prior how much payload size wise I'll insert then rockdb is better right if i want to save disk space ??

[narendra9079]

I have did a testing where I inserted 100K points with the below payload between RocksDb vs Grid Store.

Points: Dataset: fashion_Mnist_DinoV2_Embeddings
Payload: "This layered approach allows Gridstore to locate available space quickly, scaling down the work required for scans while keeping memory overhead minimal. With this system, finding storage space for new values requires scanning only a tiny fraction of the total metadata, making updates and insertions highly efficient, even in large segments."

100K points: RocksDB

Time taken for 100K upsert: 67274 ms and collection directory size 162MB and average search performance for 10 search requests=> 15ms
RAM usages: After launch 39MB, After create collection 42MB, after upsert peak ram ~263MB, after search 204.5MB (From 203.7MB to 204.5 for search request)

100K point: Grid store: (Block size: 128 Bytes, Region Block size: 8192 Bytes, Tracker size: 1MB, Page size: 32MB)

Time taken for 100K upsert: 43017 ms and collection directory size 191MB and average search performance for 10 search requests=> 19ms
RAM usages: After launch 37.6MB, After create collection 37.7MB, after upsert peak ram ~322MB, after search 382.8MB (From 235.8MB to 382.8 for search request)

I can see there is benefit in upsert latency while the search latency is more in grid store as compare to rocks DB also there is huge RAM spike while performing search operation as compared to rocks DB.

Could you please tell why is it happening??

[generall]

It is likely, that you are measuring disk cache instread of actiual RAM usage. Gridstore only keeps tracker data in ram (see rssanon), everything else mmaps, and therefore can be evicted if needed.

[generall]

also, using 10 requests as a benchmark won't give you statistical significance

[timvisee]

Note that RocksDB will be completely removed in future versions. I'd therefore recommend not to consider RocksDB at all for any deployments.

[narendra9079]

Thank you, I am more concenrned about memory because I use qdrant for edge devices so in that scenario my memory is very constaint to give in advance, in that scenario what should i use grid store or rocksdb ??