I have two fields a and b, where b has substantially higher selectivity than a.
Now, if I am only querying on both a and b (never on either field by itself), which of the following two indexes is better and why:
{a: 1, b : 1}
{b: 1, a : 1}
Explain seems to return almost identical results, but I read somewhere that you should put higher selectivity fields first. I don't know why that would make sense though.
After some extensive work to improve queries on a 150 000 000 records database I have found out the following:
not necessarily higher selectivity fields, but actually fields that are "faster" to match, being moved to the first position can increase performance drastically
I had an index composed of the following fields:
zip, address, city, first name, last name
Address is matched by an array, not string = string so it takes most time to execute and is the slowest to match. My first index that I created was: address_zip_city_last_name_first_name and the execution time for matching 1000 records against the whole DB would go for hours.
Address field actually probably has the highest selectivity on these, but since it is not being matched by a simple string equality, it takes the most time. It actually goes something like this
{ address: {$all : ["1233", "main", "avenue] }}
By changing this index to having the "faster" fields in the beginning, for example: zip_city_first_name_last_name_address the performance was much better. The same 1000 records would match in just one second instead for going for hours.
Hope this helps someone
cheers
After doing some further analysis the two indexes are in fact pretty much identical from a performance point of view.
Really if you are in a similar situation, the real consideration should be whether in the future you might be more likely to query on a alone or b alone, and put that field first in the index.
I believe the optimiser will choose the index best to use, although you can provide hints
e.g.
db.collection.find({user:u, foo:d}).hint({user:1});
see http://www.mongodb.org/display/DOCS/Optimization
Related
Considering I have search pannel that inculude multiple options like in the picture below:
I'm working with mongo and create compound index on 3-4 properties with specific order.
But when i run a different combinations of searches i see every time different order in execution plan (explain()). Sometime i see it on Collection scan (bad) , and sometime it fit right to the index (IXSCAN).
The selective fields that should handle by mongo indexes are:(brand,Types,Status,Warehouse,Carries ,Search - only by id)
My question is:
Do I have to create all combination with all fields with different order , it can be 10-20 compound indexes. Or 1-3 big Compound Index , but again it will not solve the order.
What is the best strategy to deal with big various of fields combinations.
I use same structure queries with different combinations of pairs
// Example Query.
// fields could be different every time according to user select (and order) !!
db.getCollection("orders").find({
'$and': [
{
'status': {
'$in': [
'XXX',
'YYY'
]
}
},
{
'searchId': {
'$in': [
'3859447'
]
}
},
{
'origin.brand': {
'$in': [
'aaaa',
'bbbb',
'cccc',
'ddd',
'eee',
'bundle'
]
}
},
{
'$or': [
{
'origin.carries': 'YYY'
},
{
'origin.carries': 'ZZZ'
},
{
'origin.carries': 'WWWW'
}
]
}
]
}).sort({"timestamp":1})
// My compound index is:
{status:1 ,searchId:-1,origin.brand:1, origin.carries:1 , timestamp:1}
but it only 1 combination ...it could be plenty like
a. {status:1} {b.status:1 ,searchId:-1} {c. status:1 ,searchId:-1,origin.brand:1} {d.status:1 ,searchId:-1,origin.brand:1, origin.carries:1} ........
Additionally , What will happened with Performance write/read ? , I think write will decreased over reads ...
The queries pattern are :
1.find(...) with '$and'/'$or' + sort
2.Aggregation with Match/sort
thanks
Generally, indexes are only useful if they are over a selective field. This means the number of documents that have a particular value is small relative to the overall number of documents.
What "small" means varies on the data set and the query. A 1% selectivity is pretty safe when deciding whether an index makes sense. If an particular value exists in, say, 10% of documents, performing a table scan may be more efficient than using an index over the respective field.
With that in mind, some of your fields will be selective and some will not be. For example, I suspect filtering by "OK" will not be very selective. You can eliminate non-selective fields from indexing considerations - if someone wants all orders which are "OK" with no other conditions they'll end up doing a table scan. If someone wants orders which are "OK" and have other conditions, whatever index is applicable to other conditions will be used.
Now that you are left with selective (or at least somewhat selective) fields, consider what queries are both popular and selective. For example, perhaps brand+type would be such a combination. You could add compound indexes that match popular queries which you expect to be selective.
Now, what happens if someone filters by brand only? This could be selective or not depending on the data. If you already have a compound index on brand+type, you'd leave it up to the database to determine whether a brand only query is more efficient to fulfill via the brand+type index or via a collection scan.
Continue in this manner with other popular queries and fields.
So you have subdocuments, ranged queries, and sorting by 1 field only.
It can eliminate most of the possible permutations. Assuming there are no other surprises.
D. SM already covered selectivity - you should really listen what the man says and at least upvote.
The other things to consider is the order of the fields in the compound index:
fields that have direct match like $eq
fields you sort on
fields with ranged queries: $in, $lt, $or etc
These are common rules for all b-trees. Now things that are specific to mongo:
A compound index can have no more than 1 multikey index - the index by a field in subdocuments like "origin.brand". Again I assume origins are embedded docs, so the document's shape is like this:
{
_id: ...,
status: ...,
timestamp: ....,
origin: [
{brand: ..., carries: ...},
{brand: ..., carries: ...},
{brand: ..., carries: ...}
]
}
For your query the best index would be
{
searchId: 1,
timestamp: 1,
status: 1, /** only if it is selective enough **/
"origin.carries" : 1 /** or brand, depending on data **/
}
Regarding the number of indexes - it depends on data size. Ensure all indexes fit into RAM otherwise it will be really slow.
Last but not least - indexing is not a one off job but a lifestyle. Data change over time, so do queries. If you care about performance and have finite resources you should keep an eye on the database. Check slow queries to add new indexes, collect stats from user's queries to remove unused indexes and free up some room. Basically apply common sense.
I noticed this one-year-old topic, because I am more or less struggling with a similar issue: users can request queries with an unpredictable set of the fields, which makes it near to impossible to decide (or change) how indexes should be defined.
Even worse: the user should indicate some value (or range) for the fields that make up the sharding-key, otherwise we cannot help MongoDB to limit its search in only a few shards (or chunks, for that matter).
When the user needs the liberty to search on other fields that are not necessariy the ones which make up the sharding-key, then we're stuck with a full-database search. Our dbase is some 10's of TB size...
Indexes should fit in RAM ? This can only be achieved with small databases, meaning some 100's GB max. How about my 37 TB database ? Indexes won't fit in RAM.
So I am trying out a POC inspired by the UNIX filesystem structures where we have inodes pointing to data blocks:
we have a cluster with 108 shards, each contains 100 chunks
at insert time, we take some fields of which we know they yield a good cardinality of the data, and we compute the sharding-key with those fields; the document goes into the main collection (call it "Main_col") on that computed shard, so with a certain chunk-number (equals our computed sharding-key value)
from the original document, we take a few 'crucial' fields (the list of such fields can evolve as your needs change) and store a small extra document in another collection (call these "Crucial_col_A", Crucial_col_B", etc, one for each such field): that document contains the value of this crucial field, plus an array with the chunk-number where the original full document has been stored in the 'big' collection "Main_col"; consider this as a 'pointer' to the chunk in collecton "Main_col" where this full document exists. These "Crucial_col_X" collections are sharded based on the value of the 'crucial' field.
when we insert another document that has the same value for some 'crucial' field "A", then that array in "Crucial_col_A" with chunk-numbers with be updated (with 'merge') to contain the different or same chunk number of this next full document from "Main_col"
a user can now define queries with criteria for at least one of those 'crucial' fields, plus (optional) any other criteria on other fields in the documents; the first criterium for the crucial field (say field "B") will run very quickly (because sharded on the value of "B") and return the small document from "Crucial_col_B", in which we have the array of chunk-numbers in "Main_col" where any document exists that has field "B" equal to the given criterium. Then we run a second set of parallel queries, one for each shardkey-value=chunk-number (or one per shard, to be decided) that we find in the array from before. We combine the results of those parallel subqueries, and then apply further filtering if the user gave additional criteria.
Thus this involves 2 query-steps: first in the "Crucial_col_X" collection to obtain the array with chunk-numbers where the full documents exist, and then the second query on those specific chunks in "Main_col".
The first query is done with a precise value for the 'crucial' field, so the exact shard/chunk is known, thus this query goes very fast.
The second (set of) queries are done with precise values for the sharding-keys (= the chunk numbers), so these are expected to go also very fast.
This way of working would eliminate the burden of defining many index combinations.
Let's say I have two collections, A and B, and a single document in A is related to N documents in B. For example, the schemas could look like this:
Collection A:
{id: (int),
propA1: (int),
propA2: (boolean)
}
Collection B:
{idA: (int), # id for document in Collection A
propB1: (int),
propB2: (...),
...
propBN: (...)
}
I want to return properties propB2-BN and propA2 from my API, and only return information where (for example) propA2 = true, propB6 = 42, and propB1 = propA1.
This is normally fairly simple - I query Collection B to find documents where propB6 = 42, collect the idA values from the result, query Collection A with those values, and filter the results with the Collection A documents from the query.
However, adding skip and limit parameters to this seems impossible to do while keeping the behavior users would expect. Naively applying skip and limit to the first query means that, since filtering occurs after the query, less than limit documents could be returned. Worse, in some cases no documents could be returned when there are actually still documents in the collection to be read. For example, if the limit was 10 and the first 10 Collection B documents returned pointed to a document in Collection A where propA2 = false, the function would return nothing. Then the user would assume there's nothing left to read, which may not be the case.
A slightly less naive solution is to simply check if the return count is < limit, and if so, repeat the queries until the return count = limit. The problem here is that skip/limit queries where the user would expect exclusive sets of documents returned could actually return the same documents.
I want to apply skip and limit at the mongo query level, not at the API level, because the results of querying collection B could be very large.
MapReduce and the aggregation framework appear to only work on a single collection, so they don't appear to be alternatives.
This seems like something that'd come up a lot in Mongo use - any ideas/hints would be appreciated.
Note that these posts ask similar sounding questions but don't actually address the issues raised here.
Sounds like you already have a solution (2).
You cannot optimize/skip/limit on first query, depending on search you can perhaps do it on second query.
You will need a loop around it either way, like you write.
I suppose, the .skip will always be costly for you, since you will need to get all the results and then throw them away, to simulate the skip, to give the user consistent behavior.
All the logic would have to go to your loop - unless you can match in a clever way to second query (depending on requirements).
Out of curiosity: Given the time passed, you should have a solution by now?!
I got a query like this that gets called 90% of the times:
db.xyz.find({ "ws.wz.eId" : 665 , "ws.ce1.id" : 665)
and another one like this that gets called 10% of the times:
db.xyz.find({ "ws.wz.eId" : 111 , "ws.ce2.id" : 111)
You can see that the id for the two collections in both queries are the same.
Now I'm wondering if I should just create a single index just for "ws.wz.eId" or if I should create two compound indexes: one for {"ws.wz.eId", "ws.ce.id"} and another one for {"ws.wz.eId", "ws.ce2.id"}
It seems to me that the single index is the best choice; however I might be wrong; so I would like to know if there is value in creating the compound index, or any other type.
As muratgu already pointed out, the best way to reason about performance is to stop reasoning and start measuring instead.
However, since measurements can be quite tricky, here's some theory:
You might want to consider one compound index {"ws.wz.eId", "ws.ce1.id"} because that can be used for the 90% case and, for the ten percent case, is equivalent to just having an index on ws.wz.eId.
When you do this, the first query can be matched through the index, the second query will have to find all candidates with matching ws.wz.eId first (fast, index present) and then scan-and-match all candidates to filter out those documents that don't match the ws.ce2.id criterion. Whether that is expensive or not depends on the number of documents with same ws.wz.eId that must be scanned, so this depends very much on your data.
An important factor is the selectivity of the key. For example, if there's a million documents with same ws.wz.eId and only one of those has the ws.ce2.id you're looking for, you might need the index, or want to reverse the query.
Suppose I have a compound index { a: 1, b: 1 }.
The query db.Collection.find( { b: 1 } ) doesn't use this index. The query optimizer does not appear to select this index as a candidate run.
However if you specifically hint the index, the query runs much faster and the nscan is much lower:
db.Collection.find( { b: 1 } ).hint( { a: 1, b: 1 } )
My question is, if using the index results in a faster query, why would the query optimizer ignore the index in my query on b alone?
From the page you link to on "compound index": "Compound indexes support queries on any prefix of the fields in the index." The case where an index helps on a query that is not a prefix is fairly specific, and has something to do with the distribution of values of a (I believe it does a better job as the number of possible values of a decreases). The optimal thing to do in that case is to not try using an index, because that could make things slower.
In the comments, you suggest that it shouldn't be very much slower in the worst case, but could give large improvements. Well, let's try a little testing. I built a collection with 10^6 documents, where each document i is {a: i, b: i+1}. This is, in my hypothesis, the worst case for a query on only b when using the index {a: 1, b: 1}.
For the query
db.testing.find({b: 0}).explain()
we find that it scanned 1,000,000 documents (not surprising) in about 350ms. Not bad for an unindexed query. Now, let's hint that index:
db.testing.find({b: 0}).hint("a_1_b_1").explain()
This time it only scanned 954,546 documents. I don't know enough about MongoDB indexes to explain this. However, this slightly smaller scan took about 2300ms, or 6.5x as long as the unindexed query.
So yes, a poorly-indexed query can be much worse than an unindexed one. But this doesn't completely answer your question - why doesn't the query optimizer figure this out?
The query optimizer runs different plans in parallel the first time it sees a query, and remembers the best for future queries (this is occasionally re-evaluated). But, it will only try candidate indexes - that is, those where some non-empty prefix of the index matches some portion of the query. By this standard, of course, {a: 1, b: 1} is not a candidate index for a query on just b.
I would suggest either creating a second index on {b: 1} (or at least with that prefix), or reversing the order of the one you already have (create {b: 1, a: 1} and then drop the old one).
Compound index are generally used, for prefix matched queries, or full matched ones.
Clearly your first query don't qualify. You don't need to provide a hack for this. Instead you can just hint the optimiser to use the { a : 1, b : 1 } index
db.Collection.find({ b: 1 }).hint({ a:1, b:1 })
If you have a phone book that is organized by "Last name, First name" but you only had a first name, do you think the phone book would help you find the person you were searching for?
That's what you are trying to force the optimizer to do when you have an index on a, b and you are selecting on b. It means for every value of a it needs to look and see if b matches.
There are many possible reasons why using this index may be faster than a collection scan in some circumstances. In general, it's not a candidate index and you should not use this as a solution to speeding up queries on b.
The way the current version's MongoDB query optimizer works is it tries the query with multiple query plans (all candidate indexes plus collection scan). Whichever is fastest "wins", the others are terminated and the winning plan is cached for some period of time. If you run `db.collection.find(...).explain(true) you will actually see all the "plans" it has tried. If the index is not considered a candidate then it won't be in the mixed for this phase - the only way to get the query to use it would be to explicitly "hint" it.
The query optimizer will be changing in the next major release so the above applies to the state of the world in 2.4 and earlier versions.
We have two types of high-volume queries. One looks for docs involving 5 attributes: a date (lte), a value stored in an array, a value stored in a second array, one integer (gte), and one float (gte).
The second includes these five attributes plus two more.
Should we create two compound indices, one for each query? Assume each attribute has a high cardinality.
If we do, because each query involves multiple arrays, it doesn't seem like we can create an index because of Mongo's restriction. How do people structure their Mongo databases in this case?
We're using MongoMapper.
Thanks!
Indexes for queries after the first ranges in the query the value of the additional index fields drops significantly.
Conceptually, I find it best to think of the addition fields in the index pruning ever smaller sub-trees from the query. The first range chops off a large branch, the second a smaller, the third smaller, etc. My general rule of thumb is only the first range from the query in the index is of value.
The caveat to that rule is that additional fields in the index can be useful to aid sorting returned results.
For the first query I would create a index on the two array values and then which ever of the ranges will exclude the most documents. The date field is unlikely to provide high exclusion unless you can close the range (lte and gte). The integer and float is hard to tell without knowing the domain.
If the second query's two additional attributes also use ranges in the query and do not have a significantly higher exclusion value then I would just work with the one index.
Rob.