You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@cassandra.apache.org by "ZhaoYang (JIRA)" <ji...@apache.org> on 2017/08/23 09:47:00 UTC
[jira] [Commented] (CASSANDRA-13787) RangeTombstoneMarker and
ParitionDeletion is not properly included in MV
[ https://issues.apache.org/jira/browse/CASSANDRA-13787?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16138159#comment-16138159 ]
ZhaoYang commented on CASSANDRA-13787:
--------------------------------------
The first problem is not exactly a MV bug, it's SSTableReader issue of handling multiple slices. The existing code consumes the original RangeTombstoneMarker for the first slice to generate a RangeTombstoneMarker with first slice's clustering. So there is no tombstones generated for other slices. The fix is to the ClusteringPrefix comparison, to make sure tombstones are generated probably for each slices within the range of original RangeTombstoneMarker.
The second problem is that if there is no existing live data, empty existing row is given to the ViewUpdateGenerator which may resurrect deleted cells. The fix is to include current-open-deletion into the empty existing row. So view could shadow the deleted cells.
[Draft|https://github.com/jasonstack/cassandra/commits/CASSANDRA-13787-3.0] patch base on cassandra-3.0 . This needs to be fixed in 3.0/3.11/trunk.
> RangeTombstoneMarker and ParitionDeletion is not properly included in MV
> ------------------------------------------------------------------------
>
> Key: CASSANDRA-13787
> URL: https://issues.apache.org/jira/browse/CASSANDRA-13787
> Project: Cassandra
> Issue Type: Bug
> Components: Local Write-Read Paths, Materialized Views
> Reporter: ZhaoYang
> Assignee: ZhaoYang
>
> Found two problems related to MV tombstone.
> 1. Range-tombstone-Marker being ignored after shadowing first row, subsequent base rows are not shadowed in TableViews.
> If the range tombstone was not flushed, it was used as deleted row to shadow new updates. It works correctly.
> After range tombstone was flushed, it was used as RangeTombstoneMarker and being skipped after shadowing first update. The bound of RangeTombstoneMarker seems wrong, it contained full clustering, but it should contain range or it should be multiple RangeTombstoneMarkers for multiple slices(aka. new updates)
> 2. Partition tombstone is not used when no existing live data, it will resurrect deleted cells. It was found in 11500 and included in that patch.
> In order not to make 11500 patch more complicated, I will try fix range/partition tombstone issue here.
> {code:title=Tests to reproduce}
> @Test
> public void testExistingRangeTombstoneWithFlush() throws Throwable
> {
> testExistingRangeTombstone(true);
> }
> @Test
> public void testExistingRangeTombstoneWithoutFlush() throws Throwable
> {
> testExistingRangeTombstone(false);
> }
> public void testExistingRangeTombstone(boolean flush) throws Throwable
> {
> createTable("CREATE TABLE %s (k1 int, c1 int, c2 int, v1 int, v2 int, PRIMARY KEY (k1, c1, c2))");
> execute("USE " + keyspace());
> executeNet(protocolVersion, "USE " + keyspace());
> createView("view1",
> "CREATE MATERIALIZED VIEW view1 AS SELECT * FROM %%s WHERE k1 IS NOT NULL AND c1 IS NOT NULL AND c2 IS NOT NULL PRIMARY KEY (k1, c2, c1)");
> updateView("DELETE FROM %s USING TIMESTAMP 10 WHERE k1 = 1 and c1=1");
> if (flush)
> Keyspace.open(keyspace()).getColumnFamilyStore(currentTable()).forceBlockingFlush();
> String table = KEYSPACE + "." + currentTable();
> updateView("BEGIN BATCH " +
> "INSERT INTO " + table + " (k1, c1, c2, v1, v2) VALUES (1, 0, 0, 0, 0) USING TIMESTAMP 5; " +
> "INSERT INTO " + table + " (k1, c1, c2, v1, v2) VALUES (1, 0, 1, 0, 1) USING TIMESTAMP 5; " +
> "INSERT INTO " + table + " (k1, c1, c2, v1, v2) VALUES (1, 1, 0, 1, 0) USING TIMESTAMP 5; " +
> "INSERT INTO " + table + " (k1, c1, c2, v1, v2) VALUES (1, 1, 1, 1, 1) USING TIMESTAMP 5; " +
> "INSERT INTO " + table + " (k1, c1, c2, v1, v2) VALUES (1, 1, 2, 1, 2) USING TIMESTAMP 5; " +
> "INSERT INTO " + table + " (k1, c1, c2, v1, v2) VALUES (1, 1, 3, 1, 3) USING TIMESTAMP 5; " +
> "INSERT INTO " + table + " (k1, c1, c2, v1, v2) VALUES (1, 2, 0, 2, 0) USING TIMESTAMP 5; " +
> "APPLY BATCH");
> assertRowsIgnoringOrder(execute("select * from %s"),
> row(1, 0, 0, 0, 0),
> row(1, 0, 1, 0, 1),
> row(1, 2, 0, 2, 0));
> assertRowsIgnoringOrder(execute("select k1,c1,c2,v1,v2 from view1"),
> row(1, 0, 0, 0, 0),
> row(1, 0, 1, 0, 1),
> row(1, 2, 0, 2, 0));
> }
> @Test
> public void testRangeDeletionWithFlush() throws Throwable
> {
> testExistingParitionDeletion(true);
> }
> @Test
> public void testRangeDeletionWithoutFlush() throws Throwable
> {
> testExistingParitionDeletion(false);
> }
> public void testExistingParitionDeletion(boolean flush) throws Throwable
> {
> // for partition range deletion, need to know that existing row is shadowed instead of not existed.
> createTable("CREATE TABLE %s (a int, b int, c int, d int, PRIMARY KEY (a))");
> execute("USE " + keyspace());
> executeNet(protocolVersion, "USE " + keyspace());
> createView("mv_test1",
> "CREATE MATERIALIZED VIEW %s AS SELECT * FROM %%s WHERE a IS NOT NULL AND b IS NOT NULL PRIMARY KEY (a, b)");
> Keyspace ks = Keyspace.open(keyspace());
> ks.getColumnFamilyStore("mv_test1").disableAutoCompaction();
> execute("INSERT INTO %s (a, b, c, d) VALUES (?, ?, ?, ?) using timestamp 0", 1, 1, 1, 1);
> if (flush)
> FBUtilities.waitOnFutures(ks.flush());
> assertRowsIgnoringOrder(execute("SELECT * FROM mv_test1"), row(1, 1, 1, 1));
> // remove view row
> updateView("UPDATE %s using timestamp 1 set b = null WHERE a=1");
> if (flush)
> FBUtilities.waitOnFutures(ks.flush());
> assertRowsIgnoringOrder(execute("SELECT * FROM mv_test1"));
> // remove base row, no view updated generated.
> updateView("DELETE FROM %s using timestamp 2 where a=1");
> if (flush)
> FBUtilities.waitOnFutures(ks.flush());
> assertRowsIgnoringOrder(execute("SELECT * FROM mv_test1"));
> // restor view row with b,c column. d is still tombstone
> updateView("UPDATE %s using timestamp 3 set b = 1,c = 1 where a=1"); // upsert
> if (flush)
> FBUtilities.waitOnFutures(ks.flush());
> assertRowsIgnoringOrder(execute("SELECT * FROM mv_test1"), row(1, 1, 1, null));
> }
> {code}
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@cassandra.apache.org
For additional commands, e-mail: commits-help@cassandra.apache.org