[Assistants] Add new indices for efficient querying

[mishig25]

Read this comment

[mishig25]

we have this filter (for hf.co/chat/assistants), which is

const filter: Filter<Assistant> = {
	modelId: modelId ?? { $exists: true },
	...(!createdByCurrentUser && { userCount: { $gt: 1 } }),
	...(createdByName ? { createdByName } : { featured: true }),
};
...
.sort({ userCount: -1 })

[coyotte508]

Looking at Assistant.ts, modelId is always defined.

So the filter should be like this:

const filter: Filter<Assistant> = {
	...(modelId && {modelId}),
	...(!createdByCurrentUser && { userCount: { $gt: 1 } }),
	...(createdByName ? { createdByName } : { featured: true }),
};
...
.sort({ userCount: -1 })

There is already an index on createdBy, but createdBy doesn't exist in Assistant.ts, there is instead createdByName and createdById.

IMO:

• Change the createdBy index to {createdById: 1, userCount: -1}
• Change the filter to use createdById if possible (by doing a first query to find user matching name). Otherwise add an index {createdByName: 1, userCount: -1}
• Change the featured index to {featured: 1, userCount: -1}
• Create an index on {modelId: 1, userCount: -1}. No need for a more specific index at the moment, if we add a more specific index the other key should be added before userCount, eg {modelId: 1, featured: 1, userCount: -1). It's important that the "sort" key is last.

[mishig25]

great suggestions. Will apply those.

if we add a more specific index the other key should be added before userCount, eg {modelId: 1, featured: 1, userCount: -1). It's important that the "sort" key is last.

I don't see any information the "sort" key is last on the docs here. Why so ?

[coyotte508]

Check out this video in their resources: https://learn.mongodb.com/learn/course/mongodb-indexes/lesson-4-working-with-compound-indexes-in-mongodb/learn?page=1

Basically in a compound index you want to:

• have an exact match on key1, key2, ... keyN
• sort on keyN+1

For example, if you define an index {featured: 1, userCount: -1}, and you have those documents:

{_id: "id0", featured: true, userCount: 1},
{_id: "id1", featured: true, userCount: 1},
{_id: "id2", featured: true, userCount: 2},
{_id: "id3", featured: true, userCount: 3},
{_id: "id4", featured: false, userCount: 1},
{_id: "id5", featured: false, userCount: 2},
{_id: "id6", featured: false, userCount: 3},

The index will have this internal structure:

{
  [featured=true]: {
    [userCount=3]: ["id3"],
    [userCount=2]: ["id2"],
    [userCount=1]: ["id1", "id0"]
  },
 [featured=false]: {
    [userCount=3]: ["id6"],
    [userCount=2]: ["id5"],
    [userCount=1]: ["id4"]
  },
}

So if you do the query .find({featured: true}).sort({userCount: -1}), mongodb will be able to look up the desired results very efficiently: index[featured=true].entries().flatMap(([, v] => v).

If the index was defined ({userCount: -1, featured: 1}) the internal index structure would look like this:

{
  [userCount=3]: {
    [featured=true]: ["id3"],
    [featured=false]: ["id6"],
  },
  [userCount=2]: {
    [featured=true]: ["id2"],
    [featured=false]: ["id5"],
  },
  [userCount=1]: {
    [featured=true]: ["id1", "id0"],
    [featured=false]: ["id4"],
  },
}

and MongoDB would have to examine all index keys to return the results:

index
  .entries()
  .flatMap(
    ([, v]) => v
                   .entries().filter(k => k === "featured=true")
                   .flatMap([, v] => v)
  );

It will examine all 7 indexed values instead of only the 4 returned.

This rule - putting "sort" last - is only when we have an exact match on the previous keys. Otherwise there are tradeoffs between the values being already sorted & the number of values examined.

[coyotte508]

In general https://learn.mongodb.com/ is an awesome resource to learn from - but the format is videos, not text

[nsarrazin]

damn, learned a lot here, thanks 😁

[mishig25]

@coyotte508 wdyt 2e87ab5 ?

[nsarrazin]

I don't have enough mongo knowledge to chime in so if it's good for @coyotte508 then it's good for me 😁

Do we need to manually build the indices in atlas or will they be picked up on space rebuild ?

[coyotte508]

It will automatically build

We should delete the removed indicies manually though

[mishig25]

these indices were deleted:

assistants.createIndex({ createdBy: 1 }).catch(console.error);
assistants.createIndex({ featured: 1 }).catch(console.error);