This is an automated email from the ASF dual-hosted git repository.
dataroaring pushed a commit to branch branch-2.1
in repository https://gitbox.apache.org/repos/asf/doris.git
The following commit(s) were added to refs/heads/branch-2.1 by this push:
new b489cdf8408 [opt](merge-on-write) avoid to check delete bitmap while
lookup rowkey in some situation to reduce CPU cost (#41480) (#41439)
b489cdf8408 is described below
commit b489cdf840856eb553973edde7fcd056d3ef92b1
Author: zhannngchen <48427519+zhannngc...@users.noreply.github.com>
AuthorDate: Fri Oct 11 10:15:39 2024 +0800
[opt](merge-on-write) avoid to check delete bitmap while lookup rowkey in
some situation to reduce CPU cost (#41480) (#41439)
## Proposed changes
Issue Number: close #xxx
cherry-pick #41480
---
be/src/olap/tablet.cpp | 27 +++++++++++++++++----------
be/src/olap/tablet.h | 3 ++-
2 files changed, 19 insertions(+), 11 deletions(-)
diff --git a/be/src/olap/tablet.cpp b/be/src/olap/tablet.cpp
index c1a94223eff..9f07f17ab8a 100644
--- a/be/src/olap/tablet.cpp
+++ b/be/src/olap/tablet.cpp
@@ -2822,7 +2822,7 @@ Status Tablet::lookup_row_key(const Slice& encoded_key,
TabletSchema* latest_sch
const std::vector<RowsetSharedPtr>&
specified_rowsets,
RowLocation* row_location, uint32_t version,
std::vector<std::unique_ptr<SegmentCacheHandle>>& segment_caches,
- RowsetSharedPtr* rowset, bool with_rowid) {
+ RowsetSharedPtr* rowset, bool with_rowid, bool
is_partial_update) {
SCOPED_BVAR_LATENCY(g_tablet_lookup_rowkey_latency);
size_t seq_col_length = 0;
// use the latest tablet schema to decide if the tablet has sequence
column currently
@@ -2839,6 +2839,8 @@ Status Tablet::lookup_row_key(const Slice& encoded_key,
TabletSchema* latest_sch
Slice(encoded_key.get_data(), encoded_key.get_size() -
seq_col_length - rowid_length);
RowLocation loc;
+ bool need_to_check_delete_bitmap = is_partial_update || with_seq_col;
+
for (size_t i = 0; i < specified_rowsets.size(); i++) {
auto& rs = specified_rowsets[i];
auto& segments_key_bounds =
rs->rowset_meta()->get_segments_key_bounds();
@@ -2877,15 +2879,19 @@ Status Tablet::lookup_row_key(const Slice& encoded_key,
TabletSchema* latest_sch
if (!s.ok() && !s.is<KEY_ALREADY_EXISTS>()) {
return s;
}
- if (s.ok() &&
_tablet_meta->delete_bitmap().contains_agg_without_cache(
- {loc.rowset_id, loc.segment_id, version},
loc.row_id)) {
- // if has sequence col, we continue to compare the sequence_id
of
- // all rowsets, util we find an existing key.
- if (schema->has_sequence_col()) {
- continue;
+ if (s.ok() && need_to_check_delete_bitmap) {
+ // check if the key is already mark deleted
+ if (_tablet_meta->delete_bitmap().contains_agg_without_cache(
+ {loc.rowset_id, loc.segment_id, version},
loc.row_id)) {
+ // if has sequence col, we continue to compare the
sequence_id of
+ // all rowsets, util we find an existing key.
+ if (with_seq_col) {
+ continue;
+ }
+ // The key is deleted, we need to break the loop and return
+ // KEY_NOT_FOUND.
+ break;
}
- // The key is deleted, we don't need to search for it any more.
- break;
}
// `st` is either OK or KEY_ALREADY_EXISTS now.
// for partial update, even if the key is already exists, we still
need to
@@ -3052,7 +3058,8 @@ Status Tablet::calc_segment_delete_bitmap(RowsetSharedPtr
rowset,
RowsetSharedPtr rowset_find;
auto st = lookup_row_key(key, rowset_schema.get(), true,
specified_rowsets, &loc,
- dummy_version.first - 1, segment_caches,
&rowset_find);
+ dummy_version.first - 1, segment_caches,
&rowset_find, false,
+ is_partial_update);
bool expected_st = st.ok() || st.is<KEY_NOT_FOUND>() ||
st.is<KEY_ALREADY_EXISTS>();
// It's a defensive DCHECK, we need to exclude some common errors
to avoid core-dump
// while stress test
diff --git a/be/src/olap/tablet.h b/be/src/olap/tablet.h
index fb20498355f..fdfad28a0de 100644
--- a/be/src/olap/tablet.h
+++ b/be/src/olap/tablet.h
@@ -438,7 +438,8 @@ public:
const std::vector<RowsetSharedPtr>&
specified_rowsets,
RowLocation* row_location, uint32_t version,
std::vector<std::unique_ptr<SegmentCacheHandle>>&
segment_caches,
- RowsetSharedPtr* rowset = nullptr, bool with_rowid =
true);
+ RowsetSharedPtr* rowset = nullptr, bool with_rowid =
true,
+ bool is_partial_update = false);
// Lookup a row with TupleDescriptor and fill Block
Status lookup_row_data(const Slice& encoded_key, const RowLocation&
row_location,
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@doris.apache.org
For additional commands, e-mail: commits-h...@doris.apache.org
