Mongo Notes from MongoSF 20011

MongoSF_Notes.md

MongoDB Notes

Building Web Apps with MongoDB

• MapReduce is a two step process
  • reduce function takes output of maps and collections and aggregates then gives you it in a collection
  • http://www.mongodb.org/display/DOCS/MapReduce

MongoDB Schema Design

• How to start thinking in terms of rich document modeling
• mongo makes you feel like you are denormalizing your data, it makes your data feel more object like
• object like is a huge gain of mongo
• collections is a set of documentions equivalent to a table
• NO joins in mongo, but there is embedding
• sophisticated query system, not as good as SQL, but pretty decent
• all updates are atomic and isolated
• Considerations
  • no joins
  • documents are atomic
• mongo id is a bson specific id that is given to you
• you get an automatic timestamp as well
• You can examine the query plan by using .explain()
  • http://www.mongodb.org/display/DOCS/Explain
• cool update operators such as puss, pull, pop, etc..
• The 'dot' operator
  • reach into the fields of the objects
  • http://www.mongodb.org/display/DOCS/Dot+Notation+%28Reaching+into+Objects%29
• Modify atomically
  • findAndModify allows you to find and modify atomically
• have the db conform to the application you are trying to build

MongoDB Performance Tuning by Shutterfly

• Shutterfly doesn't have any cloud based stuff they run on their own private servers
• traditional to RDBMS environments
• Data modeling matters, kind of where you start tuning
• General tuning order
  • modeling
  • statement tuning
  • instance tuning
  • hardware tuning
  • data modeling (http://www.mongodb.org/display/DOCS/MongoDB+Data+Modeling+and+Rails)
• Statement tuning
  • enable it, leave it on. it is a low overhead
  • What to look for?
    • full scans
      • nreturned vs nscanned
    • updates
      • fastmod (fastest)
      • moved (exceeds reserved space in document)
      • key updates (indexes need update)
  • explain()
    • use during development
    • use when you find bad operations in profiler
    • db.foo.find().explain()
      • index usage; nscanned vs nreturned
      • nYeilds = waiting for an operation to be completed
      • covered indexes says you can get all data by just reading the index no reason to go to the payload
      • run twice for in memory speed
• High performance writes
  • Tuning
    • read before write
    • profiler
      • tune for fastmod
    • architectural changes
      • split by collection
      • shard
• High performance reads
  • cache to disk ratio
    • try to have enough memory in system for your indexes
    • mongostat faults column
  • data locality
    • organize data for optimized I/O path. Minimize I/O per query
• Tools
  • mongostat
    • aggregate instance level information
      • faults: cache misses
      • lock%: tune updates
  • mtop
    • good picture of current session level information
  • iostat
    • how much physical I/O you are doing?
• is it faster to use a single thread for writes?
  • yes

MongoDB Shell hacks

• shell is spidermonkey
• what is it good for?
  • debugging
  • administration
  • scripting glue
  • NOT for building apps

Migrating from MySQL to MongoDB by Craigslist

• Endianness http://en.wikipedia.org/wiki/Endianness
• Shard http://en.wikipedia.org/wiki/Sharding
• SAN http://en.wikipedia.org/wiki/Storage_area_network
• Lesson: Replica Sets Rock
• Lesson: Know your data
  • mongodb is utf-8
• Lesson: Know your data size
  • 4mb in 1.6.x and 16mb in 1.8.x
• Lesson: Know some sharding
  • balancer can be your frenemy
  • initial insert rate: 8000/sec
  • http://blog.zawodny.com/2011/03/06/mongodb-pre-splitting-for-faster-data-loading-and-importing/

Performance indicators of MongoDB

• mongostat
  • like iostat
  • gives you your virtual size
  • provided by a database command called serverStatus
    • db.serverStatus();
  • profiler
    • db.setProfilingLevel(2)
      • 2 = any operations (insert, read, write) that takes longer than a certain amount of miliseconds the default is 100
  • principals for indexing
    • same as RDBMS
• Monitoring service
  • Nagios and Munin as well as MMS (Mongo Monitoring service)
• Write block percentage
  • Concurrency
    • one write OR many readers
• web-console
  • always have at port 28017 an http page for console info
• background flushing
  • 10gen tells people to RAID their EBS volumes
• connection leaks are sometimes an issue
• Network bytes in and out
  • important for read heavy applications
• Fragmentation
  • padding factor
    • you cannot manually set padding factor right now
    • dynamically calculated, the amount of space to leave when you update a new document
• Journaling
  • recommend having a second spindle just for the journal because syncing to the journal is a little expensive
• you can create a secondary index in the background
  • can take a secondary index offline and then sync it back up

MongoDB @ foursquare

• nginx, Haproxy
• mongodb and migrating off of postgres
• what we love about mongodb
  • fast
  • indexes and rich queries
  • sharding and auto-balancing
  • replication (see http://engineering.foursquare.com/2011/05/24/fun-with-mongodb-replica-sets/)
• lessons learned
  • keep working set in memory
    • keep indexes in memory
  • avoid long-running queries
  • monitor everything (per collection stats)
    • application level metrics is always good to monitor
  • use small field names for large collections

MongoDB in Ruby

• mongo gem and bson gem because bson is the native object
• bsonext gem make it a bit faster
• all ruby types map to bson types
• object ids are NOT strings
• MongoMapper recommended over MongoID. There is also Mongomatic

MongoDB in the Cloud

• You need to size your replica set as if it were the primary
• Typical MongoD should be on a large or extra large standard on demand instance on EC2
• Big MongoD should be on extra large, double extra large, quadruple extra large high-memory on-demand instance on EC2
• Small instance on EC2 is 32-bit so DO NOT use it
• ConfigD/Arbiter can run on a micro instance on EC2
• High-CPU Medium is 32-bit so DO NOT use it on EC2. High-CPU in general is just not necessary. More RAM is more important than having more CPU
• Operating Systems (Debian, Ubunti, Fedora, Redhat, FreeBSD)
  • Turn off atime
  • Raise file descriptor limits
    • cat >> /etc/security/limits.conf << EOF
      • hard nofile 65536
      • soft nofile 65536 EOF
  • Use ext4, xfs
  • DO NOT use large VM pages
  • Use RAID
    • RAID10 on MongoD
    • RAID1 on ConfigDB
• MongoD on EC2
  • LVM or MDADM
  • 64-Bit EC2 instance
  • stripping = partitions of mirrors
• MongoS on EC2
  • Runs on Application server
  • doesn't need disk, ebs volume, raid
  • 32 or 64 bit instance
• Arbiter on EC2
  • Meant to vote on elections
  • Normally need once a week
  • Do not run it on the same node as MongoD
  • 64 bit EC2 instance, micro or small is fine
• ConfigDB on EC2
  • LVM or MDADM
  • 64 bit EC2 instance micro or small is fine
• Deployment scenarios
  • 3 - Node replica set
    • 2 large MongoD in US-East one is primary and one is secondary with RAID 10
    • 1 secondary MongoD with priority = 0 (cannot become a primary) in US-West also with RAID 10
• why to find out which is the master
  • db.is_master?