MongoDB 3.0 introduces a new pluggable storage engine API and a new storage engine called WiredTiger. The engineering team behind WiredTiger team has a long and distinguished career, having architected and built Berkeley DB, now the world's most widely used embedded database. In this talk we will describe our original design goals for WiredTiger, including considerations we made for heavily threaded hardware, large on-chip caches, and SSD storage. We'll also look at some of the latch-free and non-blocking algorithms we've implemented, as well as other techniques that improve scaling, overall throughput and latency. Finally, we'll take a look at some of the features we hope to incorporate into WiredTiger and MongoDB in the future.

2. A Technical Introduction to WiredTiger
Michael Cahill
Director of Engineering (Storage), MongoDB

3. You may have seen this:

4. or this…

5. How does WiredTiger do it?

6. 6
What’s different about WiredTiger?
• Document-level concurrency
• In-memory performance
• Multi-core scalability
• Checksums
• Compression
• Durability with and without journaling

7. 7
This presentation is not…
• How to write stand-alone WiredTiger apps
• How to configure MongoDB with WiredTiger for your workload

8. WiredTiger Background

9. 9
Why create a new storage engine?
• Minimize contention between threads
– lock-free algorithms, hazard pointers
– eliminate blocking due to concurrency control
• Hotter cache and more work per I/O
– compact file formats
– compression
– big-block I/O

10. 10
MongoDB’s Storage Engine API
• Allows different storage engines to "plug-in"
– Different workloads have different performance characteristics
– mmap is not ideal for all workloads
– More flexibility
• mix storage engines on same replica set/sharded cluster
• Opportunity to integrate further (HDFS, native encrypted,
hardware optimized …)
• Great way for us to demonstrate WiredTiger’s performance

11. 11
Storage Engine Layer
Content
Repo
IoT Sensor
Backend
Ad Service
Customer
Analytics
Archive
MongoDB Query Language (MQL) + Native Drivers
MongoDB Document Data Model
MMAP V1 WT In-Memory ? ?
Supported in MongoDB 3.0 Future Possible Storage Engines
Management
Security
Example Future State
Experimental

12. WiredTiger Architecture

13. 13
WiredTiger Architecture
WiredTiger Engine
Schema &
Cursors
Python API C API Java API
Database
Files
Transactions
Page
read/ write
Logging
Column
storage
Block
management
Row
storage
Snapshots
Log Files
Cache
In-memory
consistency
Per-file
checkpoints

14. 14
Trees in cache
non-resident
child
ordinary pointer
root page
internal
page
internal
page
root page
leaf page
leaf page leaf page leaf page
1 memory
flush
read required
disk image

15. 15
Pages in cache
cache
data files
page images
on-disk
page
image
index
clean
page on-disk
page
image
indexdirty
page
updates
constructed
during read
skiplist
reconciled
during write

16. Concurrency
and
multi-core scaling

17. 17
Multiversion Concurrency Control (MVCC)
• Multiple versions of records kept in cache
• Readers see the version that was committed before the operation
started
– MongoDB “yields” turn large operations into small transactions
• Writers can create new versions concurrent with readers
• Concurrent updates to a single record cause write conflicts
– MongoDB retries with back-off

18. Transforming data during I/O

19. 19
Checksums
• A checksum is stored with every page
• WiredTiger stores the checksum with the page address (typically
in a parent page)
– Extra safety against reading an old, valid page image
• Checksums are validated during page read
– Detects filesystem corruption, random bitflips

20. 20
Compression
• WiredTiger uses snappy compression by default in MongoDB
• supported compression algorithms
– snappy [default]: good compression, low overhead
– zlib: better compression, more CPU
– none
• Indexes are compressed using prefix compression
– allows compression in memory

21. Durability
with and without
Journal

22. 22
Writing a checkpoint
1. Write the leaves
2. Write the internal pages, including the root
– the old checkpoint is still valid
3. Sync the file
4. Write the new root’s address to the metadata
– free pages from old checkpoints once the metadata is durable

23. 23
Durability without Journaling
• MMAPv1 uses a write-ahead log (journal) to guarantee
consistency
– Running with “nojournal” is unsafe
• WiredTiger doesn't have this need: no in-place updates
– Write-ahead log can be truncated at checkpoints
• Every 2GB or 60sec by default – configurable
– Updates are written to optional journal as they commit
• Not flushed on every commit by default
• Recovery rolls forward from last checkpoint
• Replication can guarantee durability

24. 24
Journal and recovery
• Optional write-ahead logging
• Only written at transaction commit
• snappy compression by default
• Group commit
• Automatic log archive / removal
• On startup, we rely on finding a consistent checkpoint in the
metadata
• Check LSNs in the metadata to figure out where to roll forward
from

25. 25
So, what’s different about WiredTiger?
• Multiversion Concurrent Control
– No lock manager
• Non-locking data structures
– Multi-core scalability
• Different representation of in-memory data vs on-disk
– Enabled checksums, compression
• Copy-on-write storage
– Durability without a journal

26. What’s next?

27. 27
What’s next for WiredTiger?
• WiredTiger btree access method is optional in 3.0
• plan to make it the default in 3.2
• tuning for (many) more workloads
• make Log Structured Merge (LSM) work well with MongoDB
– out-of-cache, write-heavy workloads
• adding encryption
• more advanced transactional semantics in the storage engine API

28. Thanks!
Questions?
Michael Cahill
michael.cahill@mongodb.com