Prevent Cosmos DB from splitting physical partition
Suppose we start off with a Cosmos DB collection consisting of a single physical partition. Is it guaranteed that the partition will not be split into multiple physical partitions as long as
- the collection size stays below the limit for physical partitions (currently 50 GB), and
- the provisioned throughput always remains <= 10,000 RU/s?
I am asking because I found Mark Brown's answer about provisioning a new partition at 60% (of what?) a bit confusing.
It isn't guaranteed.
The partition will split sometime before you get up to 50GB.
Assuming that your storage remains well below that and that you currently have exactly one partition then you can scale up to 10K without incurring partition splits though - at least as far as current behaviour of the product is.
What Mark's answer is referring to is a different heuristic at creation time compared to every other time.
If you provision a new collection with initial 10,000 manual RU you will be allocated two physical partitions.
This is because the calculation used then is CEILING(ProvisionedRU/6000)
So to get a single partition at creation time either provision it with autoscale max of <= 10K or manual throughput of <= 6K.
And then you can scale to manual 10K if desired subsequently.
When you scale up an existing container you will only get RU based partition splits if the new ProvisionedRU/Max Autoscale RU would be > 10000 * number of physical partitions that already exist.
The Portal has recently been enhanced to show where scaling up would cause a partition split - e.g. for a collection with 80 physical partitions I see that I can scale up to 800,000 "instantly" as no splits are needed to support that throughput.
