You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
storage: start with all the system ranges at bootstrap
Before this patch, a store would be bootstrapped with a single range,
and then we'd rely on the split queue to create splits at a statically
defined list of keys and between system tables.
This patch changes the bootstrapping so that the store starts up with
all these ranges and we don't rely on the split queue any more.
This simplifies thigs: less moving pieces for new clusters. Besides
that, there was a problem before since all the ranges deriving from the
original mother one were starting with expiration-based leases. Once
those expired, most of them would transition to epoch-based leases. That
transition is disruptive - clears the timestamp cache and such - and so
transactions running at those times (4.5s after cluster startup) would incur restarts.
This was a problem for tests.
Also, the fact that ranges were starting up with expiration-based leases
was also a problem for rangefeed tests since rangefeeds don't work on
those ranges (as I understand it).
As a result of this patch, a bunch of different tests now run with more
realistic stores: there are a million ways to create stores in tests and
most of them before were just getting a single range; now most get many.
Fixescockroachdb#32495Fixescockroachdb#31182Fixescockroachdb#31065
Release note: None
0 commit comments