building a scalable inbox system with mongodb and java
DESCRIPTION
Many user-facing applications present some kind of news feed/inbox system. You can think of Facebook, Twitter, or Gmail as different types of inboxes where the user can see data of interest, sorted by time, popularity, or other parameter. A scalable inbox is a difficult problem to solve: for millions of users, varied data from many sources must be sorted and presented within milliseconds. Different strategies can be used: scatter-gather, fan-out writes, and so on. This session presents an actual application developed by 10gen in Java, using MongoDB. This application is open source and is intended to show the reference implementation of several strategies to tackle this common challenge. The presentation also introduces many MongoDB concepts.TRANSCRIPT
Technical Account Manager Lead, MongoDB Inc
@antoinegirbal
Antoine Girbal
JavaOne 2013
Building a scalable inbox system with MongoDB and Java
Single Table En
Agenda
• Problem Overview
• Schema and queries
• Java Development
• Design Options – Fan out on Read– Fan out on Write– Bucketed Fan out on Write– Cached Inbox
• Discussion
Problem Overview
Let’s getSocial
Sending Messages
?
Reading my Inbox
?
Schema and Queries
Basic CRUD• Save your first document:> db.test.insert({ firstName: "Antoine", lastName: "Girbal" } )
• Find the document:> db.test.find({ firstName: "Antoine" } ){ _id: ObjectId("524495105889411fab0cdfa3"), firstName: "Antoine", lastName: "Girbal" }
• Update the document:> db.test.update({ _id: ObjectId("524495105889411fab0cdfa3") }, { x: 1, y: 2 } )
• Remove the document:> db.test.remove({ _id: ObjectId("524495105889411fab0cdfa3") })
• No schema definition or other declaration, it's easy!
The User Document{ "_id": ObjectId("519c12d53004030e5a6316d2"),
"address": { "streetAddress": "2600 Rafe Lane", "city": "Jackson", "state": "MS", "zip": 39201, "country": "US" }, "birthday": "IDODate("1980-12-26T00:00:00.000Z"), "company": "Parade of Shoes", "domain": "SanFranciscoAgency.com", "email": "[email protected]", "firstName": "Anthony", "gender": "male", "lastName": "Dacosta", "location": [ -90.183518, 32.368619 ],
…}
The User CollectionThe collection statistics:
> db.users.stats(){
"ns": "edges.users", "count": 1000000, // number of documents "size": 637864480, // size of all documents "avgObjSize": 637.86448, "storageSize": 845197312, "numExtents": 16, "nindexes": 2, "lastExtentSize": 227786752, "paddingFactor": 1.0000000000260925, // padding after documents"systemFlags": 1, "userFlags": 0, "totalIndexSize": 66070256, "indexSizes": { "_id_": 29212848, "uid_1": 36857408 }, "ok": 1
}
Queries on UsersFinding a user by email address… > db.users.find({ "email": "[email protected]" }).pretty(){ "_id": ObjectId("519c12d53004030e5a6316d2"),
…
By default will use a slow table scan… > db.users.find({ "email": "[email protected]" } ).explain(){ "cursor": "BasicCursor",
"nscannedObjects": 1000000, // 1m objects scanned"nscanned": 1000000, …
Use an index for fast performance… > db.users.ensureIndex({ "email": 1 } ) // does not do anything if index is there > db.users.find({ "email": "[email protected]" }).explain(){ "cursor": "BtreeCursor email_1", // Btree, sweet!
"nscannedObjects": 1, // document is found almost right away"nscanned": 1, …
Users Relationships• Here the follower / followee relationships
are of "many-to-many" type. It can be either stored as:
1. a list of followers in user2. a list of followees in user3. a relationship collection: "followees"4. two relationship collections: "followees" and
"followers".
• Ideal solutions:– a few million users and a 1000 followee limit:
Solution #2– no boundaries and relative scaling: Solution #3– no boundaries and max scaling: Solution #4
Relationship DataLet's look at a sample document:
> use edgesswitched to db edges> db.followees.findOne(){ "_id": ObjectId(), "user": "17052001”, "followee": "31554261”}
And the statistics:
> db.followees.stats(){ "ns": "edges.followees", "count": 1000000, "size": 64000048, "avgObjSize": 64.000048, "storageSize": 86310912, "numExtents": 10, "nindexes": 2, "lastExtentSize": 27869184, "paddingFactor": 1, "systemFlags": 1, "userFlags": 0, "totalIndexSize": 85561840, "indexSizes": {
"_id_": 32458720, "user_1_followee_1": 53103120 },
"ok": 1 }
Relationship QueriesTo find all the users that a user follows:
> db.followees.ensureIndex({ user: 1, followee: 1 }) // why not just index on user? We shall see > db.followees.find({user: "11622712"}) { "_id" : ObjectId("51641c02e4b0ef6827a34569"), "user" : "11622712", "followee" : "30432718" } … > db.followees.find({user: "11622712"}).explain() { "cursor" : "BtreeCursor user_1_followee_1", "n" : 66, "indexOnly" : false, "millis" : 0, // this is fast
Even faster if using a “covered” index:
> db.followees.find({user: "11622712"}, {followee: 1, _id: 0}).explain() { "cursor" : "BtreeCursor user_1_followee_1", "n" : 66, "nscannedObjects" : 0, "nscanned" : 66, "indexOnly" : true, // this means covered
To find all the followers of a user, we just need the opposite index::
> db.followees.ensureIndex({followee: 1, user: 1}) > db.followees.find({followee: "30313973"}, {user: 1, _id: 0})
Message DocumentThe message document: > db.messages.findOne(){ "_id": "ObjectId("519d4858e4b079162fe7eb12"), "uid": "48268973", // the author id"username": "Abiall", // why store the username?"text": "Lorem ipsum dolor sit amet, consectetur ...", "created": ISODate(2013-05-22T22:36:08.663Z"), "location": [ -95.470188, 37.366044 ], "tags": [ "gadgets" ] }
Collection statistics:
> db.messages.stats(){ "ns": "msg.messages", "count": 21440518, "size": 14184598000, "avgObjSize": 661.5790719235422, "storageSize": 15749418944, "numExtents": 27, "nindexes": 2, "lastExtentSize": 2146426864, "paddingFactor": 1, "systemFlags": 1, "userFlags": 0, "totalIndexSize": 1454289648, "indexSizes": {
"_id_": 695646784, "uid_1_created_1": 758642864 },
"ok": 1 }
Implementing the Outbox
The query is on "uid" and needs to be sorted by descending "created" time:
> db.messages.ensureIndex({ "uid": 1, "created": 1 } ) // use a compound index
> db.messages.find({ "uid": "31837072" } ).sort({ "created": -1 } ).limit(100){ "_id": ObjectId("519d626ae4b07916312e15b1") }, "uid": "31837072", "username": "Royague", "text": "Lorem ipsum dolor sit amet, consectetur adipisicing elit , sed do eiusmod tempor …", "created": ISODate("2013-05-23T00:27:22.369Z"), "location": [ "-118.296138", "33.772832" ], "tags": [ "Art" ] } …
> db.messages.find({ "uid": "31837072" }).sort({ "created": -1 }).limit(100).explain(){"cursor": "BtreeCursor uid_1_created_1 reverse", "n": 18, "nscannedObjects": 18, "nscanned": 18, "scanAndOrder": false, "millis": 0…
Java Development
Java support
• Java driver is open source, available on github and Maven.
• mongo.jar is the driver, bson.jar is a subset with BSON library only.
• Java driver is probably the most used MongoDB driver
• It receives active development by MongoDB Inc and the community
Driver Features
• CRUD
• Support for replica sets
• Connection pooling
• Distributed reads to slave servers
• BSON serializer/deserializer (lazy option)
• JSON serializer/deserializer
• GridFS
Message Storepublic class MessageStoreDAO implements MessageStore {
private Morphia morphia; private Datastore ds;
public MessageStoreDAO( MongoClient mongo ) { this.morphia = new Morphia(); this.morphia.map(DBMessage.class); this.ds = morphia.createDatastore(mongo, "messages"); this.ds.getCollection(DBMessage.class).
ensureIndex(new BasicDBObject("sender",1).append("sentAt",1) ); }
// get a messagepublic Message get(String user_id, String msg_id) { return (Message) this.ds.find(DBMessage.class) .filter("sender", user_id) .filter("_id", new ObjectId(msg_id)) .get(); }
Message Store// save a messagepublic Message save(String user_id, String message, Date date) { Message msg = new DBMessage( user_id, message, date ); ds.save( msg ); return msg; }
// find message by author sorted by descending timepublic List<Message> sentBy(String user_id) { return (List) this.ds.find(DBMessage.class) .filter("sender",user_id).order("-sentAt").limit(50).asList(); }
// find message by several authors sorted by descending timepublic List<Message> sentBy(List<String> user_ids) { return (List) this.ds.find(DBMessage.class) .field("sender").in(user_ids).order("-sentAt").limit(50).asList(); }
Graph StoreBelow uses Solution #4: both a follower and followee list
public class GraphStoreDAO implements GraphStore {
private DBCollection friends; private DBCollection followers; public GraphStoreDAO(MongoClient mongo) { this.followers = mongo.getDB("edges").getCollection("followers"); this.friends = mongo.getDB("edges").getCollection("friends"); followers.ensureIndex( new BasicDBObject("u",1).append("o",1), new BasicDBObject("unique", true)); friends.ensureIndex( new BasicDBObject("u",1).append("o",1), new BasicDBObject("unique",true)); }
// find users that are followedpublic List<String> friendsOf(String user_id) { List<String> theFriends = new ArrayList<String>(); DBCursor cursor = friends.find( new BasicDBObject("u",user_id), new BasicDBObject("_id",0).append("o",1)); while(cursor.hasNext()) theFriends.add( (String) cursor.next().get("o")); return theFriends; }
Graph Store// find followers of a userpublic List<String> followersOf(String user_id) { List<String> theFollowers = new ArrayList<String>(); DBCursor cursor = followers.find( new BasicDBObject("u",user_id), new BasicDBObject("_id",0).append("o",1)); while(cursor.hasNext()) theFollowers.add( (String) cursor.next().get("o")); return theFollowers;}
public void follow(String user_id, String toFollow) { friends.save( new BasicDBObject("u",user_id).append("o",toFollow)); followers.save( new BasicDBObject("u",toFollow).append("o",user_id));}
public void unfollow(String user_id, String toUnFollow) { friends.remove(new BasicDBObject("u", user_id).append("o", toUnFollow)); followers.remove(new BasicDBObject("u", toUnFollow).append("o", user_id));}
Design Options
4 Approaches (there are more)• Fan out on Read
• Fan out on Write
• Bucketed Fan out on Write
• Inbox Caches
Fan out on read
• Generally, not the right approach
• 1 document per message sent
• Reading an inbox is finding all messages sent by the list of people users follow
• Requires scatter-gather on sharded cluster
• Then a lot of random IO on a shard to find everything
Fan out on ReadPut the followees ids in a list:
> var fees = [] > db.followees.find({user: "11622712"})
.forEach( function(doc) { fees.push( doc.followee ) } )
Use $in and sort() and limit() to gather the inbox:
> db.messages.find({ uid: { $in: fees } }).sort({ created: -1 }).limit(100){ "_id": ObjectId("519d627ce4b07916312f0a09"), "uid": "34660390", "username": "Dingdowas" } …{ "_id": ObjectId("519d627ce4b07916312f0a10"), "uid": "34661390", "username": "John" } …{ "_id": ObjectId("519d627ce4b07916312f0a11"), "uid": "34662390", "username": "Brenda" } ……
Fan out on read – Send Message
Shard 1 Shard 2 Shard 3
Send Message
Fan out on read – Inbox Read
Shard 1 Shard 2 Shard 3
Read Inbox
Fan out on read > db.messages.find({ uid: { $in: fees } } ).sort({ created: -1 } ).limit(100).explain() {
"cursor": "BtreeCursor uid_1_created_1 multi", "isMultiKey": false, "n": 100, "nscannedObjects": 1319, "nscanned": 1384, "nscannedObjectsAllPlans": 1425, "nscannedAllPlans": 1490, "scanAndOrder": true, // it is sorting in RAM??"indexOnly": false, "nYields": 0, "nChunkSkips": 0, "millis": 31 // takes about 30ms
}
Fan out on read - sort
Fan out on write
• Tends to scale better than fan out on read
• 1 document per recipient
• Reading my inbox is just finding all of the messages with me as the recipient
• Can shard on recipient, so inbox reads hit one shard
• But still lots of random IO on the shard
Fan out on Write// Shard on “recipient” and “sent” db.shardCollection(”myapp.inbox”, { ”recipient”: 1, ”sent”: 1 } )
msg = { from: "Joe”, sent: new Date(), message: ”Hi!” }
// Send a message, write one message per followerfor( follower in followersOf( msg.from) ) {
msg.recipient = recipientdb.inbox.save(msg);
}
// Read my inbox, super easydb.inbox.find({ recipient: ”Joe” }).sort({ sent: -1 })
Fan out on write – Send Message
Shard 1 Shard 2 Shard 3
Send Message
Fan out on write– Read Inbox
Shard 1 Shard 2 Shard 3
Read Inbox
Bucketed Fan out on write• Each “inbox” document is an array of
messages
• Append a message onto “inbox” of recipient
• Bucket inbox documents so there’s not too many per document
• Can shard on recipient, so inbox reads hit one shard
• 1 or 2 documents to read the whole inbox
Bucketed Fan out on Write
// Shard on “owner / sequence”db.shardCollection(”myapp.buckets”, { ”owner”: 1, ”sequence”: 1 } )db.shardCollection(”myapp.users”, { ”user_name”: 1 } )
msg = { from: "Joe”, sent: new Date(), message: ”Hi!” }
// Send a message, have to find the right sequence documentfor( follower in followersOf( msg.from) ) { sequence = db.users.findAndModify({ query: { user_name: recipient}, update: { '$inc': { ’msg_count': 1 }}, upsert: true, new: true }).msg_count / 50;
db.buckets.update({ owner: recipient, sequence: sequence}, { $push: { ‘messages’: msg } }, { upsert: true });
}
// Read my inboxdb.buckets.find({ owner: ”Joe” }).sort({ sequence: -1 }).limit(2)
Bucketed fan out on write - Send
Shard 1 Shard 2 Shard 3
Send Message
Bucketed fan out on write - Read
Shard 1 Shard 2 Shard 3
Read Inbox
Cached inbox
• Recent messages are fast, but older messages are slower
• Store a cache of last N messages per user
• Used capped array to age out older messages
• Create cache lazily when user accesses inbox
• Only write the message if cache exists.
• Use TTL collection to time out caches for inactive users
Cached Inbox// Shard on “owner"db.shardCollection(”myapp.caches”, { ”owner”: 1 } )
// Send a message, add it to the existing caches of followersfor( follower in followersOf( msg.from) ) {
db.caches.update({ owner: recipient }, { $push: { messages: {$each: [ msg ], $sort: { ‘sent’: 1 },$slice: -50 } } } );
// Read my inboxIf( msgs = db.caches.find({ owner: ”Joe” }) ) {
// cache document existsreturn msgs;
} else {// fall back to "fan out on read" and cache itdb.caches.save({owner:’joe’, messages:[]});msgs = db.outbox.find({sender: { $in: [ followersOf( msg.from ) ] }}).sort({sent:-1}).limit(50);db.caches.update({user:’joe’}, {$push: msgs });
}
Cached Inbox – Send
Shard 1 Shard 2 Shard 3
Send Message
Cached Inbox- Read
Shard 1 Shard 2 Shard 3
Read Inbox
1
2
Cache Hit
Cache Miss
Discussion
TradeoffsFan out on Read
Fan out on Write
Bucketed Fan out on
Write
Inbox Cache
Send Message Performance
Best Single shardSingle write
GoodShard per recipientMultiple writes
WorstShard per recipientAppends (grows)
MixedDepends on how many users are in cache
Read Inbox Performance
WorstBroadcast all shardsRandom reads
GoodSingle shardRandom reads
Best Single shardSingle read
MixedRecent messages fastOlder messages are slow
Data Size Best Message stored once
WorstCopy per recipient
WorstCopy per recipient
GoodSame as FoR + size of cache
Things to consider
• Lots of recipients
• Fan out on write might become prohibitive• Consider introducing a “Group” • Make fan out asynchronous
• Very large message size
• Multiple copies of messages can be a burden• Consider single copy of message with a “pointer” per
inbox
• More writes than reads
• Fan out on read might be okay
Summary
• Multiple ways to model status updates
• Think about characteristics of your network – Number of users – Number of edges – Publish frequency – Access patterns
• Try to minimize random IO
Technical Account Manager Lead, MongoDB Inc
Antoine Girbal
JavaOne 2013
Thank You