1. MongoDB in the Cloud
Justin Smestad
@jsmestad
justin@mongomachine.com
Mongo Boulder
Jan 21, 2011

2. docs.mongomachine.com

3. Why host in the “Cloud”?

4. It makes MongoDB a commodity

5. Make scaling & deployment easy

6. The Platforms

7. Platforms: Amazon AWS
The Good
Large RAM Sizes
Community
Paired Services (SQS,
S3, etc)
Reserved Instances

8. Platforms: Amazon AWS
The Good
Large RAM Sizes
Community
Paired Services (SQS,
S3, etc)
Reserved Instances

9. Platforms: Amazon AWS
The Good
Large RAM Sizes
Community
Paired Services (SQS,
S3, etc)
Reserved Instances

10. Platforms: Amazon AWS
The Good
Large RAM Sizes
Community
Paired Services (SQS,
S3, etc)
Reserved Instances

11. Platforms: Amazon AWS
The Not So Good
Lack of small 64-bit
instances
IO Bottlenecks
EBS Volumes

12. Platforms: Amazon AWS
The Not So Good
Lack of small 64-bit
instances
IO Bottlenecks
EBS Volumes

13. Platforms: Amazon AWS
The Not So Good
Lack of small 64-bit
instances
IO Bottlenecks
EBS Volumes

14. Platforms: Rackspace
The Good
Internal Disk
Raw Performance
(CPU vs. ECPU)
64-bit instances

15. Platforms: Rackspace
The Good
Internal Disk
Raw Performance
(CPU vs. ECPU)
64-bit instances

16. Platforms: Rackspace
The Good
Internal Disk
Raw Performance
(CPU vs. ECPU)
64-bit instances

17. Platforms: Rackspace
The Not So Good
Lack of large RAM
instances *
Community
Lack of paired services

18. Platforms: Rackspace
The Not So Good
Lack of large RAM
instances *
Community
Lack of paired services

19. Platforms: Rackspace
The Not So Good
Lack of large RAM
instances *
Community
Lack of paired services

20. Platforms: Terremark
The Good
VMWare-based
Fastest Cloud Option
Fiber connected drives

21. Platforms: Terremark
The Good
VMWare-based
Fastest Cloud Option
Fiber connected drives

22. Platforms: Terremark
The Good
VMWare-based
Fastest Cloud Option
Fiber connected drives

23. Platforms: Terremark
The Not So Good
Lack of large RAM
instances *
Community
Lack of paired services

24. Platforms: Terremark
The Not So Good
Lack of large RAM
instances *
Community
Lack of paired services

25. Platforms: Terremark
The Not So Good
Lack of large RAM
instances *
Community
Lack of paired services

26. but don’t forget...
it’s not all sunshine and rainbows

27. Headaches

28. Instance Selection

29. Headaches: Instance Selection
32-bit versus 64-bit
High Memory versus High CPU
Instance Location (even inside availability zones)

30. Headaches: Instance Selection
32-bit versus 64-bit
High Memory versus High CPU
Instance Location (even inside availability zones)

31. Headaches: Instance Selection
32-bit versus 64-bit
High Memory versus High CPU
Instance Location (even inside availability zones)

32. Our Suggestion
Largest instance you can afford to buy in pairs.*
* - that will meet your dataset requirements.

33. Disk Performance

34. Headaches: Disk Performance
Ephemeral versus Elastic Block Store (EBS)
EBS Performance
Choosing a RAID type (0, 1, or 5)
Instance or EBS based AMI (machine images)

35. Headaches: Disk Performance
Ephemeral versus Elastic Block Store (EBS)
EBS Performance
Choosing a RAID type (0, 1, or 5)
Instance or EBS based AMI (machine images)

36. Headaches: Disk Performance
Ephemeral versus Elastic Block Store (EBS)
EBS Performance
Choosing a RAID type (0, 1, or 5)
Instance or EBS based AMI (machine images)

37. Headaches: Disk Performance
Ephemeral versus Elastic Block Store (EBS)
EBS Performance
Choosing a RAID type (0, 1, or 5)
Instance or EBS based AMI (machine images)

38. Our Suggestion
Use RAID-0 (if acceptable) and an EBS-based
AMI. Do not rely on EBS performance to be
constant.

39. Growth & Expansion

40. Headaches: Growth/Expansion
EBS Resizing (they must be detached)
RAID 0 & RAID 1 can’t be resized
LVM and XFS provide answers, but also add
complexity
{fsync: 1, lock: 1} is important. Also don’t close the
session!

41. Headaches: Growth/Expansion
EBS Resizing (they must be detached)
RAID 0 & RAID 1 can’t be resized
LVM and XFS provide answers, but also add
complexity
{fsync: 1, lock: 1} is important. Also don’t close the
session!

42. Headaches: Growth/Expansion
EBS Resizing (they must be detached)
RAID 0 & RAID 1 can’t be resized
LVM and XFS provide answers, but also add
complexity
{fsync: 1, lock: 1} is important. Also don’t close the
session!

43. Headaches: Growth/Expansion
EBS Resizing (they must be detached)
RAID 0 & RAID 1 can’t be resized
LVM and XFS provide answers, but also add
complexity
{fsync: 1, lock: 1} is important. Also don’t close the
session!

44. Our Suggestion
Use mdadm with LVM for RAID to enhance IO
performance while keeping the ability to expand.
This makes snapshotting volumes very hard.*
* - alternative option is using RAID-5 but you will lose some performance.

45. Disaster Recovery & Failover

46. Headaches: Disaster Recovery
Can feel a bit like “Yak Shaving”
To be good at it you will have to practice.
Knowing what can fail (ex: entire availability zones)
... this is going to take a few slides ...

47. Headaches: Disaster Recovery
Can feel a bit like “Yak Shaving”
To be good at it you will have to practice.
Knowing what can fail (ex: entire availability zones)
... this is going to take a few slides ...

48. Headaches: Disaster Recovery
Can feel a bit like “Yak Shaving”
To be good at it you will have to practice.
Knowing what can fail (ex: entire availability zones)
... this is going to take a few slides ...

49. Headaches: Disaster Recovery
Can feel a bit like “Yak Shaving”
To be good at it you will have to practice.
Knowing what can fail (ex: entire availability zones)
... this is going to take a few slides ...

50. Headaches: Disaster Recovery
EBS Reliability (better than physical, but not by much)
Snapshotting to S3 (expensive)
Replication Scenarios
Connection Strings

51. Headaches: Disaster Recovery
EBS Reliability (better than physical, but not by much)
Snapshotting to S3 (expensive)
Replication Scenarios
Connection Strings

52. Headaches: Disaster Recovery
EBS Reliability (better than physical, but not by much)
Snapshotting to S3 (expensive)
Replication Scenarios
Connection Strings

53. Headaches: Disaster Recovery
EBS Reliability (better than physical, but not by much)
Snapshotting to S3 (expensive)
Replication Scenarios
Connection Strings

54. Our Suggestion
- Practice recovery scenarios even the unlikely ones.
- Replication only gets you so far unless you go across
availability zones.
- Calculate your “acceptable downtime” and base your
infrastructure around that. If you want high availability &
recovery, your better off paying someone else.

55. Why Mongo Machine?
Moving On.

56. mongo hosted MongoDB + data management
www.mongomachine.com

57. Our Product

58. Managed SaaS
Wherever you need it.

59. mongo machine hosted MongoDB + data management
Managed MongoDB Infrastructure
Our infrastructure on AWS or Rackspace
Management & Deployment to your own infrastructure
Management Console
Instantly create new databases
Track database trends
Scale up & down on-demand

60. Management Console
Analytics data about what your DB is doing.

61. Management Console
Analytics data about what your DB is doing.

62. mongo machine hosted MongoDB + data management
Our Goals
Increase developer efficiency
Be transparent with customers
Educate our clients

63. Questions?
@jsmestad / @mongo_machine
http://docs.mongomachine.com