Deployment
- 2. Thoughts on Deployment roger@10Gen.com @rogerb
- 3. Congratulations ! Development done ? Toll ! Ready to Deploy :-)
- 4. some points to consider
- 5. Agenda • Sizing Your Hardware • memory / cpu / disk io • Software • os / filesystem • Installing MongoDB • EC2 Notes • Security • Backup • Durability • Upgrading • Monitoring • Scaling out
- 6. Sizing Hardware: Memory • Memory Mapped files • Maps DB on Filesystem to Memory • Minor Page Faults - in memory - cheap • Major Page Faults - not in memory - from disk - expensive • Indices • Part of the regular DB files • Can be completely in memory .. • Or partially • Working set should be as much in memory as possible, but
- 7. Sizing Hardware: CPU • MongoDB uses multiple cores • For working-set queries, CPU usage is minimal • Generally, faster CPU are better • Aggregation, Full Tablescans •Makes heavy use of CPU / Disk •Instead of counting / computing: • cache / precompute • Map Reduce • Currently Single threaded •Parallel in sharded environments. • This restriction may be eliminated, investigating options
- 8. Sizing Hardware: I/O • Disk I/O determines performance of non-working set queries • More Disks = Better • Improved throughput, Reduced Seek times • Raid 0 - Striping: improved write performance • Raid 1 - Mirroring: survive single disk failure • Raid 10 - both • Flash •Expensive, getting cheaper •Significantly reduced seek time, increased IO throughput • Network •It’s easy to saturate your network
- 9. IOStat iostat -‐x 2 iostat -‐w 1 disk0 disk1 disk2 cpu load average KB/t tps MB/s KB/t tps MB/s KB/t tps MB/s us sy id 1m 5m 15m 12.83 3 0.04 2.01 0 0.00 12.26 2 0.02 11 5 83 0.35 0.26 0.25 11.12 75 0.81 0.00 0 0.00 0.00 0 0.00 60 24 16 0.68 0.34 0.28 4.00 3 0.01 0.00 0 0.00 0.00 0 0.00 60 23 17 0.68 0.34 0.28 avg-cpu: %user %nice %system %iowait %steal %idle 0.00 0.00 7.96 29.85 0.50 61.69 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 sda2 0.50 4761.19 6.47 837.31 75.62 43681.59 51.86 38.38 42.33 0.46 38.41 Monitor disk transfers : > 200 -‐ 300 Mb/s on XL EC2, but your mileage may vary CPU usage > 30 % during normal operations
- 10. OS • For production: Use a 64bit OS • 32bit has 2G limit • Clients can be 32 bit • MongoDB supports (little endian only): • Linux • Windows • Solaris (joyent)
- 11. Filesystem • All data, namespace files stored in data directory • Possible to create links • Better to aggregrate IO across disks •File Allocation
- 12. Filesystem • MongoDB is filesystem-neutral: • ext3, ext4 and XFS are most used • ext4 / XFS preferred (posix_fileallocate()) • improved performance for file allocation • Support for NTFS for windows
- 13. MongoDB Version Policy • Production: run even numbers • 1.4.x, 1.6.x, 1.8.x •Development •1.5.x, 1.7.x • Critical bugs are back ported to even versions
- 14. Installing MongoDB • Installing from Source • Requires Scons, C++ compiler, Boost libraries, ... • Installing from Binaries (easiest) • curl -O http://downloads.mongodb.org/_os_/_version_ • Upgrading database • Install new version of MongoDB • Stop previous version • Start new version
- 15. EC2 Notes • Default storage instance is EXT3 • For best performance, reformat to EXT4 / XFS • Use Striping (using MDADM or LVM) aggregates I/O •This is a good thing • EC2 can experience spikes in latency • 400-600mS •This is a bad thing
- 16. More EC2 Notes • EBS snapshots can be used for backups • EBS can disappear • S3 can be used for longer term backups • Use Amazon availability zones • High Availability • Disaster Recovery
- 17. Security • Mongo supports basic security • We encourage to run mongoDB in a safe environment • Authenticates a User on a per Database basis • Start database with --auth • Admin user stored in the admin database use admin db.addUser("administrator", "password") db.auth(“administrator”, “password”) • Regular users stored in other databases use personnel db.addUser("joe", "password") db.addUser(“fred”, “password”, true)
- 18. Backup • Typically backups are driven from a slave • Eliminates impact to client / application traffic to master
- 19. Backup •Two main Strategies • mongodump / mongorestore • Filesystem • Filelock + fsync
- 20. mongodump • binary, compact object dump • each consistent object is written • not necessarily consistent from start to finish • mongorestore to restore database • database does not have to be up to restore
- 21. Filesystem Backup • MUST •fsync - flushes buffers to disk • lock - blocks writes db.runCommand({fsync:1,lock:1}) • Use file-system / LVM / storage snapshot • unlock db.$cmd.sys.unlock.findOne();
- 22. Durability What failures do you need to recover from? • Loss of a single database node? • Loss of a group of nodes?
- 23. Durability - Master only • Write acknowledged when in memory on master only
- 24. Durability - Master + Slaves • W=2 •Write acknowledged when in memory on master + slave • Will survive failure of a single node
- 25. Durability - Master + Slaves + fsync • W=n •Write acknowledged when in memory on master + slaves • Pick a “majority” of nodes • fsync in batches (since it blocking)
- 26. Slave delay • Protection against app faults • Protection against administration mistakes • Slave runs X amount of time behind
- 27. Scale out read shard1 shard2 shard3 mongos / rep_a1 rep_a2 rep_a3 config server mongos / rep_b1 rep_b2 rep_b3 config server mongos / rep_c2 rep_c2 rep_c3 config server write
- 28. Monitoring • We like Munin .. • ... but other frameworks work as well • Primary function: • Measure stats over time
- 29. Thank You :-) @rogerb
- 30. download at mongodb.org conferences, appearances, and meetups http://www.10gen.com/events Facebook | Twitter | LinkedIn http://bit.ly/mongoN @mongodb http://linkd.in/joinmongo