[DRILL-5115] Metadata Cache Pruning randomly returns wrong results at higher concurrencies

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 1.8.0, 1.9.0, 1.10.0
• Component/s: Metadata, Query Planning & Optimization
• Assignee: Unassigned
• Reporter: Rahul Kumar Challapalli
• Created: 08/Dec/16 22:20
• Updated: 30/Sep/19 11:42

Description

git.commit.id.abbrev=4312d65

When multiple queries are updating the metadata cache simultaneously the below query randomly returns wrong results.

A single run includes executing a suite of 90 tests at a concurrency of 50. I encountered a wrong data scenario in my 10th run.

Query :

select l_orderkey from l_3level where dir0=1 and ((dir1='one' and dir2 IN ('2015-7-12', '2015-7-13')) or (dir1='two' and dir2='2015-8-12'))

Wrong Result Plan (based on the profile) :

00-00    Screen : rowType = RecordType(ANY l_orderkey): rowcount = 310.0, cumulative cost = {341.0 rows, 341.0 cpu, 0.0 io, 0.0 network, 0.0 memory}, id = 205721
00-01      Project(l_orderkey=[$0]) : rowType = RecordType(ANY l_orderkey): rowcount = 310.0, cumulative cost = {310.0 rows, 310.0 cpu, 0.0 io, 0.0 network, 0.0 memory}, id = 205720
00-02        Scan(groupscan=[ParquetGroupScan [entries=[ReadEntryWithPath [path=maprfs:/drill/testdata/metadata_caching_pp/l_3level/1/one/2015-7-13/20.parquet], ReadEntryWithPath [path=maprfs:/drill/testdata/metadata_caching_pp/l_3level/1/two/2015-8-12/30.parquet], ReadEntryWithPath [path=maprfs:/drill/testdata/metadata_caching_pp/l_3level/1/one/2015-7-12/10.parquet]], selectionRoot=maprfs:/drill/testdata/metadata_caching_pp/l_3level, numFiles=3, usedMetadataFile=true, cacheFileRoot=/drill/testdata/metadata_caching_pp/l_3level, columns=[`l_orderkey`]]]) : rowType = RecordType(ANY l_orderkey): rowcount = 310.0, cumulative cost = {310.0 rows, 310.0 cpu, 0.0 io, 0.0 network, 0.0 memory}, id = 205719

Correct Result Plan (based on the profile):

00-00    Screen : rowType = RecordType(ANY l_orderkey): rowcount = 2.25, cumulative cost = {122.475 rows, 527.475 cpu, 0.0 io, 0.0 network, 0.0 memory}, id = 226849
00-01      Project(l_orderkey=[$3]) : rowType = RecordType(ANY l_orderkey): rowcount = 2.25, cumulative cost = {122.25 rows, 527.25 cpu, 0.0 io, 0.0 network, 0.0 memory}, id = 226848
00-02        SelectionVectorRemover : rowType = RecordType(ANY dir0, ANY dir1, ANY dir2, ANY l_orderkey): rowcount = 2.25, cumulative cost = {122.25 rows, 527.25 cpu, 0.0 io, 0.0 network, 0.0 memory}, id = 226847
00-03          Filter(condition=[AND(=($0, 1), OR(AND(=($1, 'one'), OR(=($2, '2015-7-12'), =($2, '2015-7-13'))), AND(=($1, 'two'), =($2, '2015-8-12'))))]) : rowType = RecordType(ANY dir0, ANY dir1, ANY dir2, ANY l_orderkey): rowcount = 2.25, cumulative cost = {120.0 rows, 525.0 cpu, 0.0 io, 0.0 network, 0.0 memory}, id = 226846
00-04            Scan(groupscan=[ParquetGroupScan [entries=[ReadEntryWithPath [path=/drill/testdata/metadata_caching_pp/l_3level/1/one/2015-7-13/20.parquet], ReadEntryWithPath [path=/drill/testdata/metadata_caching_pp/l_3level/1/two/2015-8-12/30.parquet], ReadEntryWithPath [path=/drill/testdata/metadata_caching_pp/l_3level/1/one/2015-7-12/10.parquet]], selectionRoot=/drill/testdata/metadata_caching_pp/l_3level, numFiles=3, usedMetadataFile=true, cacheFileRoot=/drill/testdata/metadata_caching_pp/l_3level/1, columns=[`dir0`, `dir1`, `dir2`, `l_orderkey`]]]) : rowType = RecordType(ANY dir0, ANY dir1, ANY dir2, ANY l_orderkey): rowcount = 60.0, cumulative cost = {60.0 rows, 240.0 cpu, 0.0 io, 0.0 network, 0.0 memory}, id = 226845

I attached the data set, log files and the query profiles. Let me know if you need anything