Random UIDs for Region Server Spread

[Nicholas Whitehead]

It seemed to me that using random UIDs was a simpler way than salting to ensure [eventual] even distribution of metrics across region servers, but my testing has not shown this yet. Is that assumption wrong, or have I not implemented correctly/with enough volume ?

Any idea ?

//Nicholas

[Thibault Godouet]

Personally I use both, and my time series are well distributed.

From memory random UID gives you a distribution of metrics, but you may still have some metrics which are way bigger than others.  Salting helps with this, and spreads the time series in the metrics around.  Combining the two should give you balanced regions.

[ManOLamancha]

I recommend both also if you expect to keep adding metrics over the life of your deployment. Particularly if you have more region servers than salt buckets. E.g. we run 20 salt buckets with 170+ region servers so that gives us a good distribution between salts and metrics.