[store] use hasCache to minimize pendingRef pool

[max-hoffman]

Expand the ChunkStore interface to let the nodeStore access recently accessed chunks. Avoid adding a child ref to the pendingRef list when already present in nbs.hasCache. For TPC-C this appears to reduce the pending ref count by another ~80%.

[max-hoffman]

#benchmark

[github-actions]

@max-hoffman workflow run: https://github.com/dolthub/dolt/actions/runs/8512757441

[coffeegoddd]

@max-hoffman DOLT

test_name	from_latency_median	to_latency_median	is_faster
tpcc-scale-factor-1	118.92	118.92	0

test_name	server_name	server_version	tps	test_name	server_name	server_version	tps	is_faster
tpcc-scale-factor-1	dolt	053e6ca	17.63	tpcc-scale-factor-1	dolt	0fa773c	17.71	0

[coffeegoddd]

@max-hoffman DOLT

comparing_percentages
100.000000 to 100.000000

version	result	total
0fa773c	ok	5937457

version	total_tests
0fa773c	5937457

correctness_percentage
100.0

[coffeegoddd]

@max-hoffman DOLT

read_tests	from_latency_median	to_latency_median	is_faster
covering_index_scan	2.97	2.91	0
groupby_scan	17.95	17.63	0
index_join	5.09	5.09	0
index_join_scan	2.26	2.22	0
index_scan	54.83	53.85	0
oltp_point_select	0.49	0.49	0
oltp_read_only	8.28	8.13	0
select_random_points	0.78	0.78	0
select_random_ranges	0.95	0.94	0
table_scan	54.83	54.83	0
types_table_scan	158.63	158.63	0

write_tests	from_latency_median	to_latency_median	is_faster
oltp_delete_insert	6.91	6.79	0
oltp_insert	3.43	3.36	0
oltp_read_write	15.83	15.55	0
oltp_update_index	3.55	3.43	0
oltp_update_non_index	3.49	3.36	0
oltp_write_only	7.84	7.56	0
types_delete_insert	7.7	7.43	0

[reltuk]

I know this was just testing perf impact...but for iterating towards landing it, I definitely don't think landing this with a new interface on the ChunkStore is the right direction.

Instead, I would think something like getAddrsCb becomes something like func(chunk []byte, func(h hash.Hash)) and then NomsBlockStore implementation uses that to actually build the hashset that it wants, and in the process filters by the concrete hasCache that was passed into tableSet.append or whatever.

[reltuk]

If perf impact of the interface change is bad, even getAddrsCb as func(chunk []byte, insertIntoThis hash.HashSet, filterByThis Haser) type Haser interface { Has(hash.Hash) bool } or func(chunk []byte, insertIntoThis hash.HashSet, filterByThis lru...) or something could be preferable...

[max-hoffman]

#benchmark

[github-actions]

@max-hoffman workflow run: https://github.com/dolthub/dolt/actions/runs/8527963737

[coffeegoddd]

@max-hoffman DOLT

test_name	from_latency_median	to_latency_median	is_faster
tpcc-scale-factor-1	121.08	118.92	0

test_name	server_name	server_version	tps	test_name	server_name	server_version	tps	is_faster
tpcc-scale-factor-1	dolt	d6aa1e6	17.91	tpcc-scale-factor-1	dolt	c832396	18.11	0

[coffeegoddd]

@max-hoffman DOLT

comparing_percentages
100.000000 to 100.000000

version	result	total
212f07f	ok	5937457

version	total_tests
212f07f	5937457

correctness_percentage
100.0

[coffeegoddd]

@max-hoffman DOLT

read_tests	from_latency_median	to_latency_median	is_faster
covering_index_scan	3.02	2.97	0
groupby_scan	17.32	17.63	0
index_join	5.09	5.09	0
index_join_scan	2.26	2.22	0
index_scan	54.83	54.83	0
oltp_point_select	0.49	0.49	0
oltp_read_only	8.13	8.13	0
select_random_points	0.78	0.78	0
select_random_ranges	0.94	0.94	0
table_scan	54.83	54.83	0
types_table_scan	161.51	158.63	0

write_tests	from_latency_median	to_latency_median	is_faster
oltp_delete_insert	6.91	6.67	0
oltp_insert	3.43	3.3	0
oltp_read_write	15.83	15.55	0
oltp_update_index	3.49	3.43	0
oltp_update_non_index	3.49	3.36	0
oltp_write_only	7.7	7.56	0
types_delete_insert	7.7	7.43	0

[reltuk]

LGTM!