asdf2014 commented on code in PR #65482:
URL: https://github.com/apache/doris/pull/65482#discussion_r3595133498
##########
be/src/runtime/query_cache/query_cache.cpp:
##########
@@ -70,4 +97,267 @@ bool QueryCache::lookup(const CacheKey& key, int64_t
version, doris::QueryCacheH
return false;
}
+bool QueryCache::lookup_any_version(const CacheKey& key, QueryCacheHandle*
handle) {
+
SCOPED_SWITCH_THREAD_MEM_TRACKER_LIMITER(ExecEnv::GetInstance()->query_cache_mem_tracker());
+ auto* lru_handle = LRUCachePolicy::lookup(key);
+ if (lru_handle) {
+ *handle = QueryCacheHandle(this, lru_handle);
+ return true;
+ }
+ return false;
+}
+
+QueryCacheInstanceDecision::~QueryCacheInstanceDecision() = default;
+
+std::unique_ptr<TabletReadSource>
QueryCacheInstanceDecision::take_delta_read_source(
+ int64_t tablet_id) {
+ std::lock_guard<std::mutex> lock(_take_lock);
+ auto it = _delta_read_sources.find(tablet_id);
+ if (it == _delta_read_sources.end()) {
+ return nullptr;
+ }
+ auto res = std::move(it->second);
+ _delta_read_sources.erase(it);
+ return res;
+}
+
+std::shared_ptr<QueryCacheInstanceDecision>
QueryCacheRuntime::get_or_make_decision(
+ const std::vector<TScanRangeParams>& scan_ranges) {
+ std::string cache_key;
+ int64_t version = 0;
+ Status st = QueryCache::build_cache_key(scan_ranges, _param, &cache_key,
&version);
+ if (!st.ok()) {
+ // No reliable cache key for this instance (e.g. FE could not align the
+ // instance to a single partition so tablets carry different versions).
+ // Degrade to an uncached scan instead of failing the query: both the
+ // scan operator and the cache source operator observe the shared MISS
+ // decision with key_valid=false, so nothing is looked up and nothing
+ // is written back. This is an expected plan shape, so log only once
+ // per fragment instead of twice per instance.
+ std::lock_guard<std::mutex> lock(_lock);
+ if (_invalid_decision == nullptr) {
+ LOG(WARNING) << "query cache degrades to uncached scan, node_id="
<< _param.node_id
+ << ", reason: " << st.to_string();
+ _invalid_decision = std::make_shared<QueryCacheInstanceDecision>();
+ }
+ return _invalid_decision;
+ }
+
+ // The lock also serializes decision making of different instances of this
+ // fragment. That is acceptable in the common case: a decision only touches
+ // tablet metadata (no data IO), so it finishes in microseconds to
+ // milliseconds. On very wide fragments (hundreds of tablets per instance)
+ // combined with tablet header lock contention (e.g. compaction meta
+ // commits) this serialization can stretch the fragment init tail; if that
+ // ever shows up in profiles, move the capture out of the lock and let the
+ // losing racer drop its duplicate decision.
+ std::lock_guard<std::mutex> lock(_lock);
+ auto it = _decisions.find(cache_key);
+ if (it != _decisions.end()) {
+ return it->second;
+ }
+
+ auto decision = std::make_shared<QueryCacheInstanceDecision>();
+ decision->key_valid = true;
+ decision->cache_key = cache_key;
+ decision->current_version = version;
+ _make_decision(scan_ranges, decision.get());
+ _decisions.emplace(std::move(cache_key), decision);
+ return decision;
+}
+
+void QueryCacheRuntime::_make_decision(const std::vector<TScanRangeParams>&
scan_ranges,
+ QueryCacheInstanceDecision* decision) {
+ decision->mode = QueryCacheInstanceDecision::Mode::MISS;
+ if (_binlog_scan) {
+ // A row-binlog scan reads a different data stream under the same plan
+ // digest: it must neither serve cached blocks nor write its output
+ // back (that would poison the cache for normal queries).
+ decision->key_valid = false;
+ return;
+ }
+ if (_param.force_refresh_query_cache) {
+ return;
+ }
+
+ QueryCacheHandle handle;
+ if (!_cache->lookup_any_version(decision->cache_key, &handle)) {
+ return;
+ }
+ // Pin the entry: as long as this decision object lives, the entry cannot
be
+ // evicted, so every operator consuming this decision sees the same blocks.
+ decision->handle = std::move(handle);
+
+ if (decision->handle.get_cache_version() == decision->current_version) {
+ decision->mode = QueryCacheInstanceDecision::Mode::HIT;
+ decision->cached_delta_count =
decision->handle.get_cache_delta_count();
+ return;
+ }
+
+ if (_try_prepare_incremental(scan_ranges, decision)) {
+ decision->mode = QueryCacheInstanceDecision::Mode::INCREMENTAL;
+ DorisMetrics::instance()->query_cache_stale_hit_total->increment(1);
+ return;
+ }
+
+ // Stale entry not reusable: drop the pin and fall back to a full scan.
+ decision->_delta_read_sources.clear();
+ decision->handle = QueryCacheHandle();
+ decision->mode = QueryCacheInstanceDecision::Mode::MISS;
+ if (!decision->incremental_fallback_reason.empty()) {
+ // Only count real fallbacks: an empty reason means incremental merge
+ // was not enabled for this query in the first place.
+
DorisMetrics::instance()->query_cache_incremental_fallback_total->increment(1);
+ }
+}
+
+bool QueryCacheRuntime::_try_prepare_incremental(const
std::vector<TScanRangeParams>& scan_ranges,
+ QueryCacheInstanceDecision*
decision) {
+ if (!(_param.__isset.allow_incremental && _param.allow_incremental)) {
+ // Not a fallback: incremental merge is simply not enabled for this
+ // query, so leave incremental_fallback_reason empty.
+ return false;
+ }
+ // Cloud tablets capture rowsets through a different (partly asynchronous)
+ // path; incremental merge only supports local storage for now.
+ if (config::is_cloud_mode()) {
+ decision->incremental_fallback_reason = "cloud mode";
+ return false;
+ }
+
+ int64_t cached_version = decision->handle.get_cache_version();
+ if (cached_version >= decision->current_version) {
+ // The entry is newer than what this replica is asked to read (e.g. the
+ // entry was filled from another replica with a newer visible version).
+ // A full scan of the requested version is the only safe answer.
+ decision->incremental_fallback_reason = "cached entry is newer";
+ return false;
+ }
+ int64_t cached_delta_count = decision->handle.get_cache_delta_count();
+ if (cached_delta_count >= config::query_cache_max_incremental_merge_count)
{
+ // Force a full recompute to compact the entry: every incremental merge
+ // appends the delta blocks to the entry, so both the entry and the
+ // upstream merge get more fragmented as deltas accumulate.
+ decision->incremental_fallback_reason = "delta count reached
compaction threshold";
+ return false;
+ }
+
+ for (const auto& scan_range : scan_ranges) {
+ int64_t tablet_id = scan_range.scan_range.palo_scan_range.tablet_id;
+ if (!_capture_tablet_delta(tablet_id, cached_version, decision)) {
+ return false;
+ }
+ }
+
+ // If the cached entry alone already exceeds the entry limits, the merged
+ // entry (which can only be larger) could never be written back. Still scan
+ // only the delta, but tell the cache source upfront so it does not clone
+ // blocks for a write back that would be discarded anyway. Such an entry
+ // stays stale until compaction merges its delta away (then the capture
+ // above fails and a full recompute takes over).
+ if (decision->handle.get_cache_total_bytes() > _param.entry_max_bytes ||
+ decision->handle.get_cache_total_rows() > _param.entry_max_rows) {
+ decision->write_back_feasible = false;
+ }
+
+ decision->cached_version = cached_version;
+ decision->cached_delta_count = cached_delta_count;
+ return true;
+}
+
+bool QueryCacheRuntime::_capture_tablet_delta(int64_t tablet_id, int64_t
cached_version,
+ QueryCacheInstanceDecision*
decision) {
+ auto tablet_res = ExecEnv::get_tablet(tablet_id);
+ if (!tablet_res) {
+ decision->incremental_fallback_reason = "tablet not found";
+ return false;
+ }
+ BaseTabletSPtr tablet = std::move(tablet_res.value());
+ // Defensive re-check of what FE promised with allow_incremental:
+ // "cached snapshot + delta rowsets == new snapshot" holds unconditionally
+ // for append-only DUP_KEYS data, and for merge-on-write UNIQUE data as
+ // long as the delta did not rewrite any row that predates the cached
+ // version (verified below once the delta rowsets are known). It never
+ // holds for merge-on-read UNIQUE data (duplicates are resolved by merging
+ // across rowsets at read time, so a delta-only scan cannot stand alone)
+ // nor for AGG tables (rows merge in the storage layer likewise).
+ const bool merge_on_write = tablet->keys_type() == KeysType::UNIQUE_KEYS &&
+ tablet->enable_unique_key_merge_on_write();
+ if (tablet->keys_type() != KeysType::DUP_KEYS && !merge_on_write) {
+ decision->incremental_fallback_reason = "keys type not append-only";
+ return false;
+ }
+
+ // quiet: on a version-graph miss (the delta merged away by compaction, an
+ // expected per-query situation) this option makes the capture API neither
+ // log nor error; it returns an EMPTY read source instead, so emptiness is
+ // checked below as the failure signal.
+ auto source_res = tablet->capture_read_source({cached_version + 1,
decision->current_version},
Review Comment:
Right, on a mid-window miss the quiet capture returns the prefix it already
walked, so the empty check alone was insufficient. Fixed in 8d4e3c39be9. The
capture result is now verified to cover the whole (cached_version,
current_version] window: the first split must start at cached_version + 1 and
the last must end at current_version. The path is contiguous by construction,
each step appends a segment starting where the previous one ended, so the two
endpoint checks pin the entire window and a partial prefix falls back to a full
recompute instead of being served and written back as the requested version.
Quiet itself stays on because a delta merged away by compaction is an expected
per-query situation that should not produce warning logs; the coverage check
gives the same safety without the noise. The new `fallback_on_partial_capture`
test builds the exact scenario above: rowsets {0,50} and {51,80} with a cached
version 50 entry and a query at version 100, expecting the "delta vers
ions not capturable" fallback instead of INCREMENTAL.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
