[jira] [Commented] (CASSANDRA-15413) Missing results on reading large frozen text map

["Tyler Codispoti (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/CASSANDRA-15413?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16976805#comment-16976805 ] 

Tyler Codispoti commented on CASSANDRA-15413:
---------------------------------------------

As a temporary workaround, we made a change to compareNextTo() in AbstractCompoundCellNameType to force using the BytesType comparator for this column. We did more changes to ensure we don't mess with any other columns, but essentially, the change boils down to:

 
{code:java}
            ....

            ByteBuffer previous = null;
            for (int i = 0; i < composite.size(); i++)
            {
                if (!hasComponent(i))
                    return nextEOC == Composite.EOC.END ? 1 : -1;

                AbstractType<?> comparator = type.subtype(i);
                ByteBuffer value1 = nextComponents[i];
                ByteBuffer value2 = composite.get(i);

                // For frozen map, do not compare each key/value. Compare the whole serilized binary
                // as what it did when writing to sstables.
                if (comparator instanceof MapType) {
                    comparator = BytesType.instance;
                }

                int cmp = comparator.compareCollectionMembers(value1, value2, previous);

                if (cmp != 0)
                    return cmp;

                previous = value1;
            }

            ....
{code}
 

This looks to resolve the issue. It seems what is happening is that, when reading Map<TEXT, TEXT>, it will compare when reading back using lexicographic order, but it is stored in binary order. When reading data back for pages after the 1st, it'll compare the last value of the previous page to the current page to see if any records can be skipped. Since the sort comparator is of the wrong type, you can easily end up in a state where it skips records in correctly.

> Missing results on reading large frozen text map
> ------------------------------------------------
>
>                 Key: CASSANDRA-15413
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-15413
>             Project: Cassandra
>          Issue Type: Bug
>          Components: Local/SSTable
>            Reporter: Tyler Codispoti
>            Assignee: Alex Petrov
>            Priority: Normal
>
> Cassandra version: 2.2.15
> I have been running into a case where, when fetching the results from a table with a frozen<map<text, text>>, if the number of results is greater than the fetch size (default 5000), we can end up with missing data.
> Side note: The table schema comes from using KairosDB, but we've isolated this issue to Cassandra itself. But it looks like this can cause problems for users of KairosDB as well.
> Repro case. Tested against fresh install of Cassandra 2.2.15.
> 1. Create table (csqlsh)
> {code:sql}
> CREATE KEYSPACE test
>   WITH REPLICATION = { 
>    'class' : 'SimpleStrategy', 
>    'replication_factor' : 1 
>   };
>   CREATE TABLE test.test (
>     name text,
>     tags frozen<map<text, text>>,
>     PRIMARY KEY (name, tags)
>   ) WITH CLUSTERING ORDER BY (tags ASC);
> {code}
> 2. Insert data (python3)
> {code:python}
> import time
> from cassandra.cluster import Cluster
> cluster = Cluster(['127.0.0.1'])
> session = cluster.connect('test')
> for i in range(0, 20000):
>     session.execute(
>         """
>         INSERT INTO test (name, tags)  
>         VALUES (%s, %s)
>         """,
>         ("test_name", {'id':str(i)})
>     )
> {code}
>  
> 3. Flush
>  
> {code:java}
> nodetools flush{code}
>  
>  
> 4. Fetch data (python3)
> {code:python}
> import time
> from cassandra.cluster import Cluster
> cluster = Cluster(['127.0.0.1'], control_connection_timeout=5000)
> session = cluster.connect('test')
> session.default_fetch_size = 5000
> session.default_timeout = 120
> count = 0
> rows = session.execute("select tags from test where name='test_name'")
> for row in rows:
>     count += 1
> print(count)
> {code}
> Result: 10111 (expected 20000)
>  
> Changing the page size changes the result count. Some quick samples:
>  
> ||default_fetch_size||count||
> |5000|10111|
> |1000|1830|
> |999|1840|
> |998|1850|
> |20000|20000|
> |100000|20000|
>  
>  
> In short, I cannot guarantee I'll get all the results back unless the page size > number of rows.
> This seems to get worse with multiple SSTables (eg nodetool flush between some of the insert batches). When using replication, the issue can get disgustingly bad - potentially giving a different result on each query.
> Interesting, if we pad the values on the tag map ("id" in this repro case) so that the insertion is in lexicographical order, there is no issue. I believe the issue also does not repro if I do not call "nodetools flush" before querying.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)

---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@cassandra.apache.org
For additional commands, e-mail: commits-help@cassandra.apache.org