phoenix-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "James Taylor (JIRA)" <j...@apache.org>
Subject [jira] [Comment Edited] (PHOENIX-418) Support approximate COUNT DISTINCT
Date Tue, 22 Aug 2017 21:00:02 GMT

    [ https://issues.apache.org/jira/browse/PHOENIX-418?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16137402#comment-16137402
] 

James Taylor edited comment on PHOENIX-418 at 8/22/17 8:59 PM:
---------------------------------------------------------------

Thanks for the revised patch, [~aertoria]. Looks very good. A couple of minor things:
- derive your test from ParallelStatsDisabledIT instead of BaseUniqueNamesOwnClusterIT and
remove the setup method which you won't need. The advantage of ParallelStatsDisabledIT tests
are that they don't need to each spin up a new mini cluster when they run so are overall test
run time stays lower.
{code}
+public class CountDistinctApproximateHyperLogLogIT extends BaseUniqueNamesOwnClusterIT {
+    @BeforeClass
+    public static void doSetup() throws Exception {
+        Map<String, String> props = Maps.newHashMapWithExpectedSize(3);
+        setUpTestDriver(new ReadOnlyProps(props.entrySet().iterator()));
+    }
+
{code}
- I think it also makes sense to have another test derived from ParallelStatsEnabledIT. This
base test class is configured to collect statistics. In this way, you can get more test coverage.
You can likely run the exact same tests, but in this case you'll have guideposts in place
(because stats will be collected). Make sure to call TestUtil.analyzeTable(connection, fullTableName)
prior to running your TABLESAMPLE queries. You'll get more rows back, since you'll have guideposts
in addition to region boundaries.
- minor nit, extra semicolon here:
{code}
+    DistinctCountHyperLogLogAggregateFunction(DistinctCountHyperLogLogAggregateFunction.class);;
{code}
- Instead of always copying the underlying byte buffer, use ByteUtil.copyKeyBytesIfNecessary(ImmutableBytesWritable
ptr) instead which only copies when necessary:
{code}
+	@Override
+	public boolean evaluate(Tuple tuple, ImmutableBytesWritable ptr) {	
+		try {
+			valueByteArray.set(hll.getBytes(), 0, hll.getBytes().length);
+			ptr.set(valueByteArray.copyBytes());
{code}


was (Author: jamestaylor):
Thanks for the revised patch, [~aertoria]. Looks very good. A couple of minor things:
- derive your test from ParallelStatsDisabledIT instead of BaseUniqueNamesOwnClusterIT and
remove the setup method which you won't need. The advantage of ParallelStatsDisabledIT tests
are that they don't need to each spin up a new mini cluster when they run so are overall test
run time stays lower.
{code}
+public class CountDistinctApproximateHyperLogLogIT extends BaseUniqueNamesOwnClusterIT {
+    @BeforeClass
+    public static void doSetup() throws Exception {
+        Map<String, String> props = Maps.newHashMapWithExpectedSize(3);
+        setUpTestDriver(new ReadOnlyProps(props.entrySet().iterator()));
+    }
+
{code}
- I think it also makes sense to have another test derived from ParallelStatsEnabledIT. This
base test class is configured to collect statistics. In this way, you can get more test coverage.
You can likely run the exact same tests, but in this case you'll have guideposts in place
(because stats will be collected). Make sure to call TestUtil.analyzeTable(connection, fullTableName)
prior to running your TABLESAMPLE queries. You'll get more rows back, since you'll have guideposts
in addition to region boundaries.
- minor nit, extra semicolon here:
{code}
+    DistinctCountHyperLogLogAggregateFunction(DistinctCountHyperLogLogAggregateFunction.class);;
{code}
- Instead of always copying the underlying byte buffer, use ByteUtil.copyKeyBytesIfNecessary(ImmutableBytesWritable
ptr) instead which only copies when necessary:
+	@Override
+	public boolean evaluate(Tuple tuple, ImmutableBytesWritable ptr) {	
+		try {
+			valueByteArray.set(hll.getBytes(), 0, hll.getBytes().length);
+			ptr.set(valueByteArray.copyBytes());
{code}

> Support approximate COUNT DISTINCT
> ----------------------------------
>
>                 Key: PHOENIX-418
>                 URL: https://issues.apache.org/jira/browse/PHOENIX-418
>             Project: Phoenix
>          Issue Type: Task
>            Reporter: James Taylor
>            Assignee: Ethan Wang
>              Labels: gsoc2016
>         Attachments: PHOENIX-418-v1.patch, PHOENIX-418-v2.patch, PHOENIX-418-v3.patch,
PHOENIX-418-v4.patch
>
>
> Support an "approximation" of count distinct to prevent having to hold on to all distinct
values (since this will not scale well when the number of distinct values is huge). The Apache
Drill folks have had some interesting discussions on this [here](http://mail-archives.apache.org/mod_mbox/incubator-drill-dev/201306.mbox/%3CJIRA.12650169.1369931282407.88049.1370645900553%40arcas%3E).
They recommend using  [Welford's method](http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance_Online_algorithm).
I'm open to having a config option that uses exact versus approximate. I don't have experience
implementing an approximate implementation, so I'm not sure how much state is required to
keep on the server and return to the client (other than realizing it'd be much less that returning
all distinct values and their counts).



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)

Mime
View raw message