RCORE-2152 Don't emit transaction log instructions for mutations on newly-created objects

[tgoyne]

The non-sync transaction logs are only used to drive notifications and notifications don't care about mutations on objects in the same commit as the objects were created it, so we don't need to emit the instructions at all. This significantly cuts the size of the transaction log for commits which are primarily inserting objects.

This does a very basic check for "newly-created" which tracks the most recently created object for each table and skips mutation instructions for that object. This handles recursively creating an object and all of its embedded objects without the overhead of tracking every single object created within a transaction, and insertion workflows will typically not return to an object after creating another object in the same table.

This requires adding an additional small amount of tracking for embedded objects, as Replication previously didn't know when new embedded objects were created.

The sample Realm which prompted this had a ton of CT history entries for updating the bootstrap store that were mildly annoying to skip over while looking for the actual bootstrap application instructions, so I made the bootstrap store not emit CT history. I don't think this is a meaningful optimization outside of very narrow circumstances (e.g. if the server decides to send us a whole bunch of single-object changesets).

realm-trawler turned out to be broken for sync Realms, so I fixed that.

The existing object-store notification tests do a pretty good job of validating that instructions are emitted when they need to be, so the new tests focus on validating that they aren't when they shouldn't, which isn't testable at the object-store level (as the point of this is to stop emitting instructions which the object-store transaction log handler was just discarding).

[coveralls-official]

Pull Request Test Coverage Report for Build thomas.goyne_391

Details

• 358 of 362 (98.9%) changed or added relevant lines in 6 files are covered.
• 74 unchanged lines in 15 files lost coverage.
• Overall coverage increased (+0.03%) to 90.876%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
src/realm/sync/noinst/pending_bootstrap_store.cpp	32	36	88.89%

Files with Coverage Reduction	New Missed Lines	%
src/realm/query_engine.hpp	1	94.11%
src/realm/util/serializer.cpp	1	90.43%
test/test_query2.cpp	1	98.73%
src/realm/replication.cpp	2	80.89%
src/realm/util/file.cpp	2	78.73%
test/test_shared.cpp	2	96.74%
test/fuzz_group.cpp	3	50.72%
src/realm/sync/instruction_applier.cpp	5	68.95%
src/realm/sync/noinst/server/server_history.cpp	5	63.44%
src/realm/sync/noinst/server/server.cpp	6	73.74%

Totals
Change from base Build 2379:	0.03%
Covered Lines:	214850
Relevant Lines:	236420

---

💛 - Coveralls

[tgoyne]

We appear to currently not have any benchmarks which exercise this very well; everything which measures insertion performance only has one column and this becomes more beneficial the more columns there are. The single column case is ~5% faster.

[danieltabacaru]

Nice work!