[
https://issues.apache.org/jira/browse/CASSANDRA-9540?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14573093#comment-14573093
]
Mathijs Vogelzang commented on CASSANDRA-9540:
----------------------------------------------
And one final observation from our testing: it seems the bug happens as soon as
a cell value exceeds 64kb in size. (My earlier comment was wrong, there were
144,000 2-byte hex chars in the json file, so the cell size is roughly 70kb
instead of 144kb).
This bug is very difficult for us to work around (we're using the IN query as a
replacement for the former thrift slice query to get specific column values),
as we currently can not trust cassandra to return values that we wrote into it
earlier, depending on whether our queried columns are present or not. If
there's any workaround how we can query a set of columns using CQL that works,
please let me know.
> Cql query doesn't return right information when using IN on columns for some
> keys
> ---------------------------------------------------------------------------------
>
> Key: CASSANDRA-9540
> URL: https://issues.apache.org/jira/browse/CASSANDRA-9540
> Project: Cassandra
> Issue Type: Bug
> Components: API
> Environment: Cassandra 2.1.5
> Reporter: Mathijs Vogelzang
> Assignee: Carl Yeksigian
> Fix For: 2.1.x
>
>
> We are investigating a weird issue where one of our clients doesn't get data
> on his dashboard. It seems Cassandra is not returning data for a particular
> key ("brokenkey" from now on).
> Some background:
> We have a row where we store a "metadata" column and data in columns
> "bucket/0", "bucket/1", "bucket/2", etc. Depending on the date selection of
> the UI, we know that we only need to retrieve bucket/0, bucket/0 and bucket/1
> etc. (we always need to retrieve "metadata").
> A typical query may look like this (using SELECT column1 to just show what is
> returned, normally we would of course do SELECT value):
> {noformat}
> cqlsh:AppBrain> select blobAsText(column1) from "GroupedSeries" where
> key=textAsBlob('install/workingkey');
> blobAsText(column1)
> ---------------------
> bucket/0
> metadata
> (2 rows)
> cqlsh:AppBrain> select blobAsText(column1) from "GroupedSeries" where
> key=textAsBlob('install/brokenkey');
> blobAsText(column1)
> ---------------------
> bucket/0
> metadata
> (2 rows)
> {noformat}
> These two queries work as expected, and return the information that we
> actually stored.
> However, when we "filter" for certain columns, the brokenkey starts behaving
> very weird:
> {noformat}
> cqlsh:AppBrain> select blobAsText(column1) from "GroupedSeries" where
> key=textAsBlob('install/workingkey') and column1 IN
> (textAsBlob('metadata'),textAsBlob('bucket/0'),textAsBlob('bucket/1'),textAsBlob('bucket/2'));
> blobAsText(column1)
> ---------------------
> bucket/0
> metadata
> (2 rows)
> cqlsh:AppBrain> select blobAsText(column1) from "GroupedSeries" where
> key=textAsBlob('install/workingkey') and column1 IN
> (textAsBlob('metadata'),textAsBlob('bucket/0'),textAsBlob('bucket/1'),textAsBlob('bucket/2'),textAsBlob('asdfasdfasdf'));
> blobAsText(column1)
> ---------------------
> bucket/0
> metadata
> (2 rows)
> *** As expected, querying for more information doesn't really matter for the
> working key ***
> cqlsh:AppBrain> select blobAsText(column1) from "GroupedSeries" where
> key=textAsBlob('install/brokenkey') and column1 IN
> (textAsBlob('metadata'),textAsBlob('bucket/0'),textAsBlob('bucket/1'),textAsBlob('bucket/2'));
> blobAsText(column1)
> ---------------------
> bucket/0
> (1 rows)
> *** Cassandra stops giving us the metadata column when asking for a few more
> columns! ***
> cqlsh:AppBrain> select blobAsText(column1) from "GroupedSeries" where
> key=textAsBlob('install/brokenkey') and column1 IN
> (textAsBlob('metadata'),textAsBlob('bucket/0'),textAsBlob('bucket/1'),textAsBlob('bucket/2'),textAsBlob('asdfasdfasdf'));
> key | column1 | value
> -----+---------+-------
> (0 rows)
> *** Adding the bogus column name even makes it return nothing from this row
> anymore! ***
> {noformat}
> There are at least two rows that malfunction like this in our table (which is
> quite old already and has gone through a bunch of Cassandra upgrades). I've
> upgraded our whole cluster to 2.1.5 (we were on 2.1.2 when I discovered this
> problem) and compacted, repaired and scrubbed this column family, which
> hasn't helped.
> Our table structure is:
> {noformat}
> cqlsh:AppBrain> describe table "GroupedSeries";
> CREATE TABLE "AppBrain"."GroupedSeries" (
> key blob,
> column1 blob,
> value blob,
> PRIMARY KEY (key, column1)
> ) WITH COMPACT STORAGE
> AND CLUSTERING ORDER BY (column1 ASC)
> AND caching = '{"keys":"ALL", "rows_per_partition":"NONE"}'
> AND comment = ''
> AND compaction = {'min_threshold': '4', 'class':
> 'org.apache.cassandra.db.compaction.SizeTieredCompactionStrategy',
> 'max_threshold': '32'}
> AND compression = {'sstable_compression':
> 'org.apache.cassandra.io.compress.LZ4Compressor'}
> AND dclocal_read_repair_chance = 0.1
> AND default_time_to_live = 0
> AND gc_grace_seconds = 864000
> AND max_index_interval = 2048
> AND memtable_flush_period_in_ms = 0
> AND min_index_interval = 128
> AND read_repair_chance = 1.0
> AND speculative_retry = 'NONE';
> {noformat}
> Let me know if I can give more information that may be helpful.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
