Andy shows a table the CMU group assembled for a survey paper comparing MVCC implementations across production systems 1:08:29.
Version storage across systems
| System | Version storage | Chain order | Notes |
|---|---|---|---|
| PostgreSQL | Append-only | Oldest → newest | The standout outlier; HOT updates avoid index churn ~40% of the time 1:08:29 56:17 1:25:03 |
| MS SQL Server Hekaton | Append-only | — | Append-only chosen because it was grafted onto SQL Server to add in-memory storage 1:09:00 |
| MySQL / InnoDB | Delta (rollback segment) | Newest → oldest | Secondary indexes store primary keys, so updates never touch them 40:10 57:09 |
| Oracle | Delta | — | Its "serializable" level is really snapshot isolation 26:48 29:35 |
| MS SQL Server (classic) | Time-travel | Newest → oldest | Retrofit onto an originally single-version 2PL engine 39:04 39:32 |
| SAP HANA | Time-travel | Oldest → newest | GC/compaction promotes the latest version into the main table 39:32 |
| HyPer / Hyrise | Delta | — | Hyrise abandoned append-only and switched to deltas 1:09:00 |