MongoDb: how to create the right (composite) index for data with many searchable fields

Question

UPDATE: I need to add that the point of this question is to allow me to define schemas for Json Rest Stores. The user can search by any one key, or several keys. So, I cannot easily predict what the users will search by -- it could be 1, 2, 5 fields (this is especially true for data-rich fields like people, bookings, etc.)

Imagine that I have an index as such:

{ "item": 1, "location": 1, "stock": 1 }

Following the MongoDb manual on indexes:

MongoDB can use this index to support queries that include:

• the item field,
• the item field and the location field,
• the item field and the location field and the stock field, or
• only the item and stock fields; however, this index would be less efficient than an index on only item and stock.

MongoDB cannot use this index to support queries that include:

• only the location field,
• only the stock field, or
• only the location and stock fields.

Now, suppose I have a schema with exactly these fields:

• item: String
• location: String
• stock: String
• qty: number

And imagine I want to make sure every query is indeed indexed. I would do:

For item:

• item, location, stock, qty
• item, location, qty, stock
• item, stock, qty, location
• item, stock, location, qty
• item, qty, location, stock
• item, qty, stock, location

For location:

• ...you know the gist

Now... this seems a little insane. If you have a database where you have TEN searchable fields, this becomes clearly unworkable as the number of indexes grows exponentially.

Am I missing something? My idea was to define a schema, define which fields were searchable, and write a function that makes up all of the needed indexes regardless of what fields were present and what fields weren't. However, I am thinking about it, and... well, I must be missing something.

Am I?

Answer 1

I will try to explain what does this mean by example. The indexes based on B-tree is not something mongodb specific. In contrast it is rather common concept.

So when you create an index - you show the database an easier way to find something. But this index is stored somewhere with a pointer pointing to a location of the original document. This information is ordered and you might look at it as binary tree which has a really nice property: the search is reduced from O(n) (linear scan) to O(log(n)). Which is much much faster because each time we trim our space in half (potentially we can reduce the time from 10^6 to 20 lookups). For example we have a big collection with field {a : some int, b: 'some other things'} and if we index it by a, we end up with another data structure which is sorted by a. It looks this way (by this I do not mean that it is another collection, this is just for demonstration):

{a : 1, pointer: to the field with a = 1}, // if a is the smallest number in the starting collection
...
{a : 999, pointer: to the field with a = 990} // assuming that 999 is the biggest field

So right now we are searching for a field a = 18. Instead of going one by one through all elements we take something in the middle and if it is bigger then 18, then we are dividing the lower part in half and checking the element there. We continue till we will find a = 18. Then we look at the pointer and knowing it we extract the original field.

The situation with compound index is similar (instead of ordering by one element we order by many). For example you have a collection:

{ "item": 5, "location": 1, "stock": 3, 'a lot of other fields' }  // was stored at position 5 on the disk
{ "item": 1, "location": 3, "stock": 1, 'a lot of other fields' }  // position 1 on the disk
{ "item": 2, "location": 5, "stock": 7, 'a lot of other fields' }  // position 3 on the disk
... huge amount of other data
{ "item": 1, "location": 1, "stock": 1, 'a lot of other fields' }  // position 9 on the disk
{ "item": 1, "location": 1, "stock": 2, 'a lot of other fields' }  // position 7 on the disk

and want an index { "item": 1, "location": 1, "stock": 1 }. The lookup table would look like this (one more time - this is not another collection, this is just for demonstration):

{ "item": 1, "location": 1, "stock": 1, pointer = 9 }
{ "item": 1, "location": 1, "stock": 2, pointer = 7 }
{ "item": 1, "location": 3, "stock": 1, pointer = 1 }
{ "item": 2, "location": 5, "stock": 7, pointer = 3 }
.. huge amount of other data (but not necessarily here. If item would be one it would be somewhere next to items 1)
{ "item": 5, "location": 1, "stock": 3, pointer = 5 }

See that here everything is basically sorted by item, then by location and then by pointer. The same way as with a single index we do not need to scan everything. If we have a query which looks for item = 2, location = 5 and stock = 7 we can quickly identify where documents with item = 2 are and then the same way quickly identify where among these items item with location 5 and so on.

And right now an interesting part. Also we created just one index (although this is a compound index, it is still one index) we can use it to quickly find the element

• only by the item. Really all we need to do is only the first step. So there is no point to create another index {location : 1} because it is already covered by compound index.
• also we can quickly find only by item and by location (we need only 2 steps).

Cool 1 index but helps us in three different ways. But wait a minute: what if we want to find by item and stock. Oh it looks like we can speed up this query as well. We can in log(n) find all elements with specific item and ... here we have to stop - magic has finished. We need to iterate through all of them. But still pretty good.

But may it can help us with other queries. Lets look at a query by location which looks like was already ordered. But if you will look at it - you see that this is a mess. One in the beginning and then one in the end. It can not help you at all.

I hope this clarifies few things:

• why indexes are good (reduce time from O(n) to potentially O(log(n))
• why compound indexes can help with some queries nonetheless we have not created an index on that particular field and help with some other queries.
• what indexes are covered by compound index
• why indexes can harm (it creates additional datastructure which should be maintained)

And this should tell another valid thing: index is not a silver bullet. You can not speed up all your queries, so it sound silly to think that by creating indexes on all fields EVERYTHING would be super fast.

Answer 2

What are your real query patterns? It's very unlikely that you would need to create all of these possible index combinations. I also doubt that including qty in the index would be of much use. Do you need to search for things where qty == 4 independent of location and item type?

An index doesn't need to identify every single record, it just needs to be specific enough to make any final scan small. Given an item code or a stock value are there really that many locations that you'd also need to index on them?

I suspect in this case an index on item, an index on location and and index on stock would be sufficient to answer most likely queries with sufficient speed. (But we'd need to know more about what these field names mean and what the count and distribution of values is within them).

Use explain with your queries and you can see how well they are performing. Add indices as necessary, don't create every possible ordering.