Add payload-size-aware large-value zone eviction for EloqKV (ObjectCcMap)

[Copilot]

EloqKV holds large values that are expensive to reload from disk. The standard LRU eviction policy does not account for payload size, causing large-value pages to drift to the eviction head at the same rate as cheap small-value entries.

Design

The LRU list is divided into two zones at a stable pointer boundary:

head ← [small-value pages] ← lru_large_value_zone_head_ ← [large-value pages] ← tail

• Small-value pages insert before lru_large_value_zone_head_ — they cycle through the LRU normally.
• Large-value pages (PayloadSize() > txservice_large_value_threshold) insert before tail_ccp_ — they always stay in the recent half regardless of how many small-value accesses occur.
• Eviction scan (head → tail) exhausts small-value pages before reaching the large-value zone. Large-value pages can still be evicted under extreme memory pressure once the SV zone is empty.
• No zone-size cap: memory accounting relies on mimalloc heap, not page count, so a fixed-ratio limit is not meaningful. In the extreme all-large-value case, SV pages are evicted first and that is accepted.

Key changes

CcShard (cc_shard.h / cc_shard.cpp)

• Add LruPage *lru_large_value_zone_head_ (init &tail_ccp_).
• UpdateLruList: insert point is &tail_ccp_ for LV pages, lru_large_value_zone_head_ for SV pages; advance boundary when the first LV page is inserted.
• DetachLru / ReplaceLru: advance/update lru_large_value_zone_head_ when the boundary page is removed/replaced.

LruPage (cc_entry.h)

• Add bool has_large_value_{false} — marks a page as belonging to the LV zone; set once, never cleared.

CcMap / ObjectCcMap (cc_map.h, object_cc_map.h)

• virtual bool IsLargeValueZoneEnabled() const { return false; } — default off for all maps.
• ObjectCcMap overrides to return txservice_large_value_threshold > 0 — policy is EloqKV-only; RangeCcMap / CatalogCcMap are unaffected.

TemplateCcMap (template_cc_map.h)

• MaybeMarkAndRezoneAsLargeValue(LruPage*, size_t) private helper: sets has_large_value_ and calls UpdateLruList to eagerly move the page into the LV zone on first detection.

ObjectCcMap (object_cc_map.h) — 6 eager re-zone call sites

• Execute(ApplyCc) — apply_and_commit_ path after payload commit
• Execute(PostWriteCc) — after commit when commit_ts > 0
• Execute(UploadBatchCc) — after PassInCurrentPayload for Normal records
• Execute(KeyObjectStandbyForwardCc) — after commands committed to payload
• Execute(ReplayLogCc) — after EmplaceAndCommitBufferedTxnCommand
• BackFill — after DeserializeCurrentPayload for Normal records

CcPageCleanGuardWithoutKickoutCc (cc_page_clean_guard.h) — lazy fallback

• On each scan, sets has_large_value_ and signals re-zone for any LV entry not yet detected by the eager path (safety net only).

tx_service_common.h

• inline size_t txservice_large_value_threshold = 0 — threshold in bytes; 0 disables the policy entirely.

tx_service.h

• Config key large_value_threshold parsed into txservice_large_value_threshold.

Here are some reminders before you submit the pull request

• Add tests for the change
• Document changes
• Reference the link of issue using fixes eloqdb/tx_service#issue_id
• Reference the link of RFC if exists
• Pass ./mtr --suite=mono_main,mono_multi,mono_basic

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[xiexiaoy]

这个算法存在两个问题：

shard.access_counter_是一个shard level的值，某个热点key就可以使它变得很大。用户不知道怎么配置txservice_large_value_eviction_age，也不想配置。

large value容易沉淀到LRU list的头部，但是因为它的age小于txservice_large_value_eviction_age，所以不会被踢掉。这造成了LRU 扫描的时候做了一些无效的工作，扫描了不能被踢的large value。

[Copilot]

Fixed in commit 55925be.

Both problems are resolved by replacing the absolute txservice_large_value_eviction_age threshold with a self-calibrating relative LRU-position check:

uint64_t now        = cc_shard_->AccessCounter();
uint64_t page_age   = now - page_->last_access_ts_;
uint64_t total_span = now - cc_shard_->LruOldestTs();
if (page_age * 2 < total_span)   // page in recent half → protect
    return {false, false};
// page in old half → allow eviction (fall through)

Problem 1 (unconfigurable threshold): txservice_large_value_eviction_age is removed entirely. The check compares a page's age against the total span of the LRU list, so it self-calibrates regardless of how fast access_counter_ grows. No user configuration is needed.

Problem 2 (cold large values blocking the scan): A cold large-value page near the LRU head has page_age ≈ total_span, so page_age * 2 >= total_span — it is immediately eligible for eviction rather than wasting scan work.