OLTP is Totally Different

by

Michael Stonebraker

The Old Way The New WayThe Old Way The New Way

storage Disk Main memory

Caching Anti-caching

ACID Record level locking

MVCC, TS ordering, …

Data logging Command logging

Replication Active-passive Active-active

Products

Old WayOracle, DB2, SQLServer, Postgres, MySQL …

New WayHekaton, Hana, MemSQL, VoltDB, SQLFire,…

Main Memory Storage

1 Tbyte costs (maybe) $25K and declining

rapidlyIf your data doesn’t fit then wait a couple of

yearsNVRAM will push this way out later this decade

Anti-Caching (VLDB ‘14)

Main memory format for dataWhen memory fills, gather cold tuples and write

to an archive (in main memory format)When a transaction has a “miss”, abort it but

continue with “fake processing” to find all the

absent dataGet and “pin” the needed dataReschedule transaction when all needed data in

main memoryNumbers from H-Store implementation

Some Data From Nirmesh Malvaiya

Implemented Aries in VoltDBCompared against the VoltDB scheme

Asynchronous checkpoints

Command logging

Some Data From Nirmesh Malvaiya

1.5 X run-time performance gain1.5 X penalty at recovery time

Almost all OLTP applications demand HAOnly run recovery for cluster-wide failures

E.g. power outage

9

Concurrency Control

MVCC popular (Hana, Hekaton) Time stamp order popular (H-Store/VoltDB) Lightweight combinations of time stamp order

and dynamic locking (Calvin, Dora) I don’t know anybody who is doing normal

dynamic locking It’s too slow!!!!

How to Implement HA

Active-Passive

As in the traditional wisdom

Active-Active

Send update transactions to all copies

Each executes transaction logic

Requires deterministic scheduling!!!

Net-Net on OLTP

Main memory DBMS

With anti-caching

And command logging

Deterministic concurrency controlHA via active-active

Has nothing to do with the traditional wisdom!!!