[seekdb][tx] Reduce tx data memtable initial hash buckets

[hnwyllmm]

Task Description

The tx data memtable hash map starts with 65,536 buckets. Each bucket header is 16 bytes, so an active memtable reserves 1 MiB for the bucket array even when the transaction count is low.

Solution Description

Reduce ObTxDataHashMap::MIN_BUCKETS_CNT from 65,536 to 16,384. This lowers the initial bucket array from 1 MiB to 256 KiB while retaining the existing adaptive expansion and shrink logic.

Passed Regressions

• Built both the 65,536 baseline and 16,384 patched release binaries successfully.
• Ran four 300-second ABBA sysbench oltp_update_non_index rounds on a 16-vCPU host, with 8 CPUs assigned to seekdb and 8 CPUs assigned to sysbench.
• Weighted throughput: 13,410.67 TPS for 65,536 buckets and 13,448.07 TPS for 16,384 buckets (+0.279%, within noise); all rounds completed with zero errors.
• The adjacent idle PSS comparison saved 718 KiB, close to the expected deterministic saving of 768 KiB per active hash map.
• git diff --check and git diff --cached --check passed.

Upgrade Compatibility

Compatible. This only changes the initial in-memory bucket allocation and does not modify persisted data, file formats, SQL behavior, or public APIs.

Other Information

This MR contains one commit: 6da09f27dbe reduce tx data memtable initial bucket count. The unrelated untracked file 2 is not included.

Release Note

[hnwyllmm]

The mapping Dima issue is about adjusting the memtable_arena parameter.

[hnwyllmm]

Local idle-memory A/B result

I re-ran the idle-memory comparison on the current host because the previous cloud host is no longer available.

Method

• Built the 65,536-bucket baseline and 16,384-bucket patched release binaries from commit 6da09f27dbe; the source difference is only MIN_BUCKETS_CNT and its comment.
• Reused the same existing base directory, port, and 2 GiB configuration.
• Waited 180 seconds without client workload and sampled /proc in four ABBA rounds.
• Because process-level RSS/PSS had multi-MiB background allocation noise, used dump chunk all to isolate the affected MEMTABLE_ARENA allocations.

Allocator dump

• 65,536 buckets: 1,048,608 + 1,048,608 + 7,936 = 2,105,152 B
• 16,384 buckets: 1,048,608 + 262,176 + 7,936 = 1,318,720 B
• Observed idle saving: 786,432 B = 768 KiB = 0.75 MiB

The changed object is the newly created ACTIVE tx-data memtable bucket array. The MINI memtable restored from the same base directory and the 7,936-byte arena object remained unchanged, so the entire 768 KiB delta is directly attributable to this change. This also agrees with the earlier adjacent idle-PSS observation of 718 KiB.

Binary SHA256

• 65,536: a631c4a472c435cd7bc4e99ad672355cf976f257502a691d681fa8d4aba8a726
• 16,384: 442dcb42fbf9856697fa96e5f43bae6897ae21f6db193088ac2c77cfd239d4ba