Details
-
Improvement
-
Status: Resolved
-
Low
-
Resolution: Fixed
-
None
-
None
Description
When deciding to initiate read repair on a certain column, in the special case that there are different column values with the same timestamp (this situation can occur if values were 'simultaneously' inserted by several clients), the behavior should be deterministic (tie-breaker). This prevents ongoing back-and-forward read repair when one node repairs towards value X, and another one repairs towards value Y.
I propose to deploy the following rule: if we are in doubt whether to repair a column with timestamp T (because two values X and Y are present within the cluster, both at timestamp T), then we always repair towards X if some_func(X)<some_func(Y). some_func() could be md5() or just plain byte-by-byte comparison (maybe the latter is better because it will never result in equality).
The major use case for this improvement is to enable Cassandra as a means to generate cluste-wide unique identifiers. This could be implemented like this:
- client wants to grab <ID>, and inserts a column with key <ID>, value <some_random_stuff>, timestamp <ID>
- client reads the just inserted column with consistency level ALL or QUORUM
- if the read value (<some_random_stuff>) matches the value originally inserted by the client, the <ID> is unique for this client
In this algorithm, collisions are prevented by the fact that if to clients try to aquire the same <ID> at the same time, the proposed tie-breaking will take care that exactly one client wins after the first read repair.