Tablet splitting: Load-based strategy

[ttyusupov]

Jira Link: DB-2587
Parent ticket: "Tablet splitting" #1004.

Depends on #1461.

[rkarthik007]

Came across a real world scenario where tablet splitting based on the load could have been useful. Documenting that here, so we can build this feature appropriately.

Use-case details:

• We were using colocated tables - this ask applies equally to colocation as to splitting tablets. The reason to use colocated tables - the use case has about 70K DB objects (yes, 70,000 tables + indexes + etc) which would have resulted in too many tables.
• Identified 40 tables in the are colocated tablet that contribute heavily to the total IOPS. The top 3 tables account for 100s of millions of ops/sec.
• Needed to move these out of co-location group, which should help immensely because the iops can now be handled by multiple vCPUs across various machines.
• The next step would likely be to split the tablet, which need to sustain a very high throughput.

Overall, we need to be able to perform tablet splits automatically based on such iops usage criteria.