Notes

Two strategies for implementing deadlock prevention used in 2-phase locking

  1. Wound wait - Force the lock to release from another transaction
  2. Wait die - Wait for lock to be released or die https://stackoverflow.com/questions/32794142/what-is-the-difference-between-wait-die-and-wound-wait-deadlock-prevention-a

Distributed Transactions

Two problems with distributed transactions

  1. Write ahead logging needs to happen on each shard
  2. There needs to be a flag on each shard indicating that its in the commit phase (because otherwise locks would be wounded)

This is solved using 2-phase commit

  1. Prepare for commit (checks if data is actually written and locks were acquired)
  2. Commit

2-phase commit is inherently not a fault tolerant protocol and that is the reason it isn’t used so much. But, in Spanner they replicated the coordinator making it fault tolerant.

Spanner

Spanner doesn’t replicate locks because of performance issues(imagine replicating locks across 100s of servers which are in different datacenters).

Leader change issues

To solve issues of leader changes, spanner makes use of checking whether leader has changed(using epoch/term numbers for leaders)