Wrap-up
- The catalog — metadata the DBMS keeps about its own data — is deferred to the next class
1:20:30 - Log-structured storage and index-organized storage are alternatives to tuple-oriented heap storage, each with pros and cons
1:20:30 - For write-heavy workloads, log-structured storage wins because it maximizes sequential I/O: whole SSTables are written at once, on flush and on compaction. Index-organized and heap storage produce mostly random writes
1:20:52 - The storage manager's abstraction is not truly airtight: once transactions and recovery enter the picture, metadata has to be exposed both upward and downward across the layers
1:20:52
Next class
- Storage models: flip the mental model from rows (all attributes of a tuple contiguous) to columns, and examine the pros and cons of column stores
1:21:27 - Also pending from this lecture: compression schemes and the null-representation results for column stores
1:18:381:15:19
Deferred to later lectures
| Topic | When |
|---|---|
| Buffer pool / memory management (Project 1) | Next lecture block 0:50 |
| Indexes and B+Trees | ~2 weeks later 41:09 |
| Bloom filters | A few weeks later 37:05 |
| Multi-versioning (MVCC) | Later in the semester 20:36 |
| Logging and recovery | Later in the semester 34:40 |
| Specialized hardware | Later in the semester 1:09:13 |