Good Schema Design and Why It Matters

A talk I gave at DjangoCon Europe 2014.

1. Andrew Godwin @andrewgodwin GOOD SCHEMA DESIGN WHY IT MATTERS and
2. Andrew Godwin Core Developer Senior Engineer Author & Maintainer
3. Schemas Explicit & Implicit
4. Explicit PostgreSQL MySQL Oracle SQLite CouchDB MongoDB Redis ZODB Implicit
5. Explicit Schema ID int Name text Weight uint 1 2 3 Alice Bob Charles 76 84 65 Implicit Schema { "id": 342, "name": "David", "weight": 44, }
6. Explicit Schema Normalised or semi normalised structure JOINs to retrieve related data Implicit Schema Embedded structure Related data retrieved naturally with object
7. Silent Failure { "id": 342, "name": "David", "weight": 74, } { "id": 342, "name": "Ellie", "weight": "85kg", } { "id": 342, "nom": "Frankie", "weight": 77, } { "id": 342, "name": "Frankie", "weight": -67, }
8. Schemas inform Storage
9. PostgreSQL
10. Adding NULLable columns: instant But must be at end of table
11. CREATE INDEX CONCURRENTLY Slower, and only one at a time
12. Constraints after column addition This is more general advice
13. MySQL Locks whole table Rewrites entire storage No DDL transactions
14. Oracle / MSSQL / etc. Look into their strengths
15. Changing the Schema Databases aren't code...
16. You can't put your database in a VCS You can put your schema in a VCS But your data won't always survive.
17. Django Migrations Codified schema change format
19. Migrations aren't enough You can't automate away a social problem!
20. What if we got rid of the schema? That pesky, pesky schema.
21. The Nesting Problem { "id": 123, "name": "Andrew", "friends": [ {"id": 456, "name": "David"}, {"id": 789, "name": "Mazz"}, ], "likes": [ {"id": 22, "liker": {"id": 789, "name", "Mazz"}}, ], }
22. You don't have to use a document DB (like CouchDB, MongoDB, etc.)
23. Schemaless Columns ID int Name text Weight uint Data json 1 Alice 76 { "nickname": "Al", "bgcolor": "#ff0033" }
24. But that must be slower... Right?
25. Comparison (never trust benchmarks) Loading 1.2 million records PostgreSQL MongoDB 76 sec 8 min Sequential scan PostgreSQL MongoDB 980 ms 980 ms Index scan (Postgres GINhash) PostgreSQL MongoDB 0.7 ms 1 ms
26. Load Shapes
27. Read-heavy Write-heavy Large size
28. Read-heavy Write-heavy Large size Wikipedia TV show page Minecraft Forums Amazon Glacier Eventbrite Logging
29. Read-heavy Write-heavy Large size Offline storage Append formats In-memory cache Many indexes Fewer indexes
30. Your load changes over time Scaling is not just a flat multiplier
31. General Advice Write heavy → Fewer indexes Read heavy → Denormalise Keep large data away from read/write heavy data Blob stores/filesystems are DBs too
32. Lessons They're near the end so you remember them.
33. Re-evaulate as you grow Different things matter at different sizes
34. Adding NULL columns is great Always prefer this if nothing else
35. You'll need more than one DBMS But don't use too many, you'll be swamped
36. Indexes aren't free You pay the price at write/restore time
37. Relational DBs are flexible They can do a lot more than JOINing normalised tables
38. Thanks! Andrew Godwin @andrewgodwin eventbrite.com/jobs are hiring: