Since a new instance can start on few seconds, can we consider that neon architecture is de facto high availability ? Will a new instance start in any available zone in case of failure ?
1 Like
Hey!
By default, we run compute instance in the same AZ with the pageserver. We’re doing this because of latencies (0.2msec vs. 1 ms). And this difference makes sense for query performance (especially for random access patterns) if the data is not cached in compute instance.
However, if one AZ is unavailable, we will spin up compute in another. In that case, it’s better to sacrifice performance over reliability.
Hey,
Do you have some experience now of the average delay when switching after an AZ failure ?