I'm trying to insert ~800 million records into MongoDB using PyMongo on a macbook air 1.7GHz i7 with no multi-threading, the documents are structured as below:
Records I'm reading are the following tuple:
(user_id,imp_date,imp_creative,imp_pid,geo_id)
I'm creating my own _id field based on the user_id in the file I'm reading from.
{_id:user_id,
'imp_date':[array of dates],
'imp_creative':[array of numeric ids],
'imp_pid':[array of numeric ids],
'geo_id':numeric id}
I'm using an upsert with $push to append date, creative id, and pid for the corresponding arrays
self.collection.update({'_id':uid},
{"$push":{'imp_date':<datevalue>,
'imp_creative':<creative_id>,
'imp_pid':<pid>}},safe=True,upsert=True)
I'm using an upsert with $set to overwrite the geographic location (only care about most recent.
self.collection.update({'_id':uid},
{"$set":{'geo_id':<geo id>}},safe=True,upsert=True)
I'm only writing about 1,500 records per second (8,000 if I set safe=False). My question is: what can I do to speed this up further (ideally 20k/second or faster)?
Ideas I can't find a definitive recommendation on:
-Using multiple threads to insert data
-Sharding
-Padding arrays (my arrays grow very slowly, each document array will have an average length of ~4 at the end of the file)
-Turning journaling off
Apologies if I've left out any required information, this is my first post.
1- You could add an index to speed it up, and index would help you to find the documents faster although the inserts would be slower (you have to update the index as well). If the improvement in the retrieving phase compensates the extra time to update the index depends on how many records you have in the collections, how many indexes you have and how complicated those indexes are.
However, in your case you are only querying with the _id so there's no much more you can do with indexes.
2- Are you using two consecutive updates? I mean, one for the $set and one for the $push?
If that's true, then you should definetelly use just one:
self.collection.update({'_id':uid},
{"$push":{'imp_date':<datevalue>,
'imp_creative':<creative_id>,
'imp_pid':<pid>},
"$set":{'geo_id':<geo id>}},
safe=True,upsert=True)
3- The update operation is an atomic operation which might locks other queries. If the document you are about to update is not already in RAM but it is in the disk, mongo will have to first fetch it from the disk and then update it. If you do a find operation first (which doesn't block as it's a read-only operation) the document will be in RAM for sure so the update operation (the locking one) will be faster:
self.collection.findOne({'_id':uid})
self.collection.update({'_id':uid},
{"$push":{'imp_date':<datevalue>,
'imp_creative':<creative_id>,
'imp_pid':<pid>},
"$set":{'geo_id':<geo id>}},
safe=True,upsert=True)
4-If your documents don't grow too much as you have said, it won't be necessary to bother about padding factor and reallocation issues. Furthermore, in some recent versions (can't remember if it was since 2.2 or 2.4) collections are created with the powerOfTwo option enabled by default.
Related
I'm using MongoDB(version 5.0), with somewhat large data and indexes.
Data Storage size is around 2.5TB and indexes(8 indexes) size is also sums up to 1TB.
Since I'm pouring those data into a new mongodb and establishing new indexes, so it's taking proximately 1~2 weeks. (Estimated, still pouring)
But, with service updates, it's highly alike that new column will have to be added into documents. (or modify exist column)
And, i think i will update them with bulkWrite and update.
Here come's the question, if the new column have nothing to do with indexes, will it be possible for update operation to work without accessing indexes?
For example, if old document, with indexes are set only on _id and memberNo
{
_id: 's9023902',
memberNo: '20219210',
purchasedMurchantNo: 'M2937'
}
is update into below (column purchasedMurchantNo updated and returned added)
{
_id: 's9023902',
memberNo: '20219210',
purchasedMurchantNo: 'NEW_FORMAT_NO_2937',
isReturned: false
}
My guess is that, it won't change the structure of index, and since mongodb will also know it by the index and query structure, so there will be no need to access index structure.
I'm asking, because with indexes, insert operations have been dramatically slowed down and cpu usage have also been dramatically increased, since this mongodb is not in service right now so high cpu usage rate is not an issue, but in the point when we update our mongodb, it will be in service, and it will be an issue.
Question in a sentense: In certain circumstances, can UpdateMany work in a fast speed, as if there is no index
Thank you in advance for brilliant answer given for not so brilliantly asked question.
I would like to understand which of the below queries would be faster, while doing updates, in mongo db? I want to update few thousands of records at one stretch.
Accumulating the object ids of those records and firing them using $in or using bulk update?
Using one or two fields in the collection which are common for those few thousand records - akin to "where" in sql and firing an update using those fields. These fields might or might not be indexed.
I know that query will be much smaller in the 2nd case as every single "_id" (oid) is not accumulated. Does accumulating _ids and using those to update documents offer any practical performance advantages?
Does accumulating _ids and using those to update documents offer any practical performance advantages?
Yes because MongoDB will certainly use the _id index (idhack).
In the second method - as you observed - you can't tell whether or not an index will be used for a certain field.
So the answer will be: it depends.
If your collection has million of documents or more, and / or the number of search fields is quite large you should prefer the first search method. Especially if the id list size is not small and / or the id values are adjacent.
If your collection is pretty small and you can tolerate a full scan you may prefer the second approach.
In any case, you should testify both methods using explain().
As far as I understood, nscannedObjects entry in the explain() method means the number of documents that MongoDB needed to go to find in the disk.
My question is: when this value is 0, what this actually mean besides the explanation above? Does MongoDB keep a cache with some documents stored there?
nscannedObjects=0 means that there was no fetching or filtering to satisfy your query, the query was resolved solely based on indexes. So for example if you were to query for {_id:10} and there were no matching documents you would get nscannedObjects=0.
It has nothing to do with the data being in memory, there is no such distinction with the query plan.
Note that in MongoDB 3.0 and later nscanned and nscannedObjects are now called totalKeysExamined and totalDocsExamined, which is a little more self-explanatory.
Mongo is a document database, which means that it can interpret the structure of the stored documents (unlike for example key-value stores).
One particular advantage of that approach is that you can build indices on the documents in the database.
Index is a data structure (usually a variant of b-tree), which allows for fast searching of documents basing on some of their attributes (for example id (!= _id) or some other distinctive feature). These are usually stored in memory, allowing very fast access to them.
When you search for documents basing on indexed attributes (let's say id > 50), then mongo doesn't need to fetch the document from memory/disk/whatever - it can see which documents match the criteria basing solely on the index (note that fetching something from disk is several orders of magnitude slower than memory lookup, even with no cache). The only time it actually goes to the disk is when you need to fetch the document for further processing (and which is not covered by the statistic you cited).
Indices are crucial to achieve high performance, but also have drawbacks (for example rarely used index can slow down inserts and not be worth it - after each insertion the index has to be updated).
I have have a Python application that is iteratively going through every document in a MongoDB (3.0.2) collection (typically between 10K and 1M documents), and adding new fields (probably doubling/tripling the number of fields in the document).
My initial thought was that I would use upsert the entire of the revised documents (using pyMongo) - now I'm questioning that:
Given that the revised documents are significantly bigger should I be inserting only the new fields, or just replacing the document?
Also, is it better to perform a write to the collection on a document by document basis or in bulk?
this is actually a great question that can be solved a few different ways depending on how you are managing your data.
if you are upserting additional fields does this mean your data is appending additional fields at a later point in time with the only changes being the addition of the additional fields? if so you could set the ttl on your documents so that the old ones drop off over time. keep in mind that if you do this you will want to set an index that sorts your results by descending _id so that the most recent additions are selected before the older ones.
the benefit of this of doing it this way is that your are continually writing data as opposed to seeking and updating data so it is faster.
in regards to upserts vs bulk inserts. bulk inserts are always faster than upserts since bulk upserting requires you to find the original document first.
Given that the revised documents are significantly bigger should I be inserting only the new fields, or just replacing the document?
you really need to understand your data fully to determine what is best but if only change to the data is additional fields or changes that only need to be considered from that point forward then bulk inserting and setting a ttl on your older data is the better method from the stand point of write operations as opposed to seek, find and update. when using this method you will want to db.document.find_one() as opposed to db.document.find() so that only your current record is returned.
Also, is it better to perform a write to the collection on a document by document basis or in bulk?
bulk inserts will be faster than inserting each one sequentially.
I have a collection of over 70 million documents. Whenever I add new documents in batches (lets say 2K), the insert operation is really slow. I suspect that is because, the mongo engine is comparing the _id's of all the new documents with all the 70 million to find out any _id duplicate entries. Since the _id based index is disk-resident, it'll make the code a lot slow.
Is there anyway to avoid this. I just want mongo to take new documents and insert it as they are, without doing this check. Is it even possible?
Diagnosing "Slow" Performance
Your question includes a number of leading assumptions about how MongoDB works. I'll address those below, but I'd advise you to try to understand any performance issues based on facts such as database metrics (i.e. serverStatus, mongostat, mongotop), system resource monitoring, and information in the MongoDB log on slow queries. Metrics need to be monitored over time so you can identify what is "normal" for your deployment, so I would strongly recommend using a MongoDB-specific monitoring tool such as MMS Monitoring.
A few interesting presentations that provide very relevant background material for performance troubleshooting and debugging are:
William Zola: The (Only) Three Reasons for Slow MongoDB Performance
Aska Kamsky: Diagnostics and Debugging with MongoDB
Improving efficiency of inserts
Aside from understanding where your actual performance challenges lie and tuning your deployment, you could also improve efficiency of inserts by:
removing any unused or redundant secondary indexes on this collection
using the Bulk API to insert documents in batches
Assessing Assumptions
Whenever I add new documents in batches (lets say 2K), the insert operation is really slow. I suspect that is because, the mongo engine is comparing the _id's of all the new documents with all the 70 million to find out any _id duplicate entries. Since the _id based index is disk-resident, it'll make the code a lot slow.
If a collection has 70 million entries, that does not mean that an index lookup involves 70 million comparisons. The indexed values are stored in B-trees which allow for a small number of efficient comparisons. The exact number will depend on the depth of the tree and how your indexes are built and the value you're looking up .. but will be on the order of 10s (not millions) of comparisons.
If you're really curious about the internals, there are some experimental storage & index stats you can enable in a development environment: Storage-viz: Storage Visualizers and Commands for MongoDB.
Since the _id based index is disk-resident, it'll make the code a lot slow.
MongoDB loads your working set (portion of data & index entries recently accessed) into available memory.
If you are able to create your ids in an approximately ascending order (for example, the generated ObjectIds) then all the updates will occur at the right side of the B-tree and your working set will be much smaller (FAQ: "Must my working set fit in RAM").
Yes, I can let mongo use the _id for itself, but I don't want to waste a perfectly good index for it. Moreover, even if I let mongo generate _id for itself won't it need to compare still for duplicate key errors?
A unique _id is required for all documents in MongoDB. The default ObjectId is generated based on a formula that should ensure uniqueness (i.e. there is an extremely low chance of returning a duplicate key exception, so your application will not get duplicate key exceptions and have to retry with a new _id).
If you have a better candidate for the unique _id in your documents, then feel free to use this field (or collection of fields) instead of relying on the generated _id. Note that the _id is immutable, so you shouldn't use any fields that you might want to modify later.