radu created FLINK-6249:
---------------------------
Summary: Distinct Aggregates for OVER window
Key: FLINK-6249
URL: https://issues.apache.org/jira/browse/FLINK-6249
Project: Flink
Issue Type: New Feature
Components: Table API & SQL
Affects Versions: 1.3.0
Reporter: radu
Time target: ProcTime/EventTime
SQL targeted query examples:
----------------------------
Q1. Boundaries are expressed in windows and meant for the elements to be
aggregated
Q1.1. `SELECT SUM( DISTINCT b) OVER (ORDER BY procTime() ROWS BETWEEN 2
PRECEDING AND CURRENT ROW) FROM stream1`
Q1.2. `SELECT SUM( DISTINCT b) OVER (ORDER BY procTime() RANGE BETWEEN
INTERVAL '1' HOUR PRECEDING AND CURRENT ROW) FROM stream1`
Q1.3. `SELECT SUM( DISTINCT b) OVER (ORDER BY rowTime() ROWS BETWEEN 2
PRECEDING AND CURRENT ROW) FROM stream1`
Q1.4. `SELECT SUM( DISTINCT b) OVER (ORDER BY rowTime() RANGE BETWEEN INTERVAL
'1' HOUR PRECEDING AND CURRENT ROW) FROM stream1`
General comments:
- DISTINCT operation makes sense only within the context of windows or
some bounded defined structures. Otherwise the operation would keep
an infinite amount of data to ensure uniqueness and would not
trigger for certain functions (e.g. aggregates)
- We can consider as a sub-JIRA issue the implementation of DISTINCT
for UNBOUND sliding windows. However, there would be no control over
the data structure to keep seen data (to check it is not re-process). -> This
needs to be decided if we want to support it (to create appropriate JIRA issues)
=> We will open sub-JIRA issues to extend the current functionality of
aggregates for the DISTINCT CASE (Q1.{1-4}). (This is the main target of this
JIRA)
=> Aggregations over distinct elements without any boundary (i.e.
within SELECT clause) do not make sense just as aggregations do not
make sense without groupings or windows.
Other similar query support
------------
Q2. Boundaries are expressed in GROUP BY clause and distinct is applied for the
elements of the aggregate(s)
`SELECT MIN( DISTINCT rowtime), prodID FROM stream1 GROUP BY FLOOR(procTime()
TO HOUR)`
=> We need to decide if we aim to support for this release distinct aggregates
for the group by (Q2). If so sub-JIRA issues need to be created. We can follow
the same design/implementation.
=> We can consider as a sub-JIRA issue the implementation of DISTINCT
for select clauses. However, there is no control over the growing
size of the data structure and it will unavoidably crash the memory.
Q3. Distinct is applied to the collection of outputs to be selected.
`SELECT STREAM DISTINCT procTime(), prodId FROM stream1 GROUP BY
FLOOR(procTime() TO DAY)`
Description:
------------
The DISTINCT operator requires processing the elements to ensure
uniqueness. Either that the operation is used for SELECT ALL distinct
elements or for applying typical aggregation functions over a set of
elements, there is a prior need of forming a collection of elements.
This brings the need of using windows or grouping methods. Therefore the
distinct function will be implemented within windows. Depending on the
type of window definition there are several options:
- Main Scope: If distinct is applied as in Q1 example for window aggregations
than either we extend the implementation with distinct aggregates (less
prefered) or extend the sliding window aggregates implementation in the
processFunction with distinctinction identification support (prefered). The
later option is prefered because a query can carry multiple aggregates
including multiple aggregates that have the distinct key word set up.
Implementing the distinction between elements in the process function avoid the
need to multiply the data structure to mark what what was seen across multiple
aggregates. It also makes the implementation more robust and resilient as we cn
keep the data structure for marking the seen elements in a state (mapstate).
- If distinct is applied as in Q2 example on group elements than
either we define a new implementation if selection is general or
extend the current implementation of grouped aggregates with
distinct group aggregates
- If distinct is applied as in Q3 example for the select all elements,
then a new implementation needs to be defined. This would work over
a specific window and within the window function the uniqueness of
the results to be processed will be done.
Functionality example
---------------------
We exemplify below the functionality of the IN/Exists when working with
streams.
`Q1: SELECT STREAM DISTINCT b FROM stream1 GROUP BY FLOOR(PROCTIME TO HOUR) `
`Q2: SELECT COUNT(DISTINCT b) FROM stream1 GROUP BY FLOOR(PROCTIME() TO HOUR)
`
`Q3: SELECT sum(DISTINCT a) OVER (ORDER BY procTime() ROWS BETWEEN 2
PRECEDING AND CURRENT ROW) FROM stream1`
<style type="text/css">
</style>
<table class="tg">
<tr>
<th class="tg-9hbo">Proctime</th>
<th class="tg-9hbo">IngestionTime(Event)</th>
<th class="tg-9hbo">Stream1</th>
<th class="tg-9hbo">Q1</th>
<th class="tg-9hbo">Q2</th>
<th class="tg-9hbo">Q3</th>
</tr>
<tr>
<td class="tg-yw4l"></td>
<td class="tg-yw4l">10:00:01</td>
<td class="tg-yw4l">(ab, 1)</td>
<td class="tg-yw4l"></td>
<td class="tg-yw4l"></td>
<td class="tg-yw4l">1</td>
</tr>
<tr>
<td class="tg-yw4l"></td>
<td class="tg-yw4l">10:05:00</td>
<td class="tg-yw4l">(aa, 2)</td>
<td class="tg-yw4l"></td>
<td class="tg-yw4l"></td>
<td class="tg-yw4l">3</td>
</tr>
<tr>
<td class="tg-yw4l">10-11</td>
<td class="tg-yw4l">11:00:00</td>
<td class="tg-yw4l"></td>
<td class="tg-yw4l">ab,aa</td>
<td class="tg-yw4l">2</td>
<td class="tg-yw4l"></td>
</tr>
<tr>
<td class="tg-yw4l"></td>
<td class="tg-yw4l">11:03:00</td>
<td class="tg-yw4l">(aa,2)</td>
<td class="tg-yw4l"></td>
<td class="tg-yw4l"></td>
<td class="tg-yw4l">3</td>
</tr>
<tr>
<td class="tg-yw4l"></td>
<td class="tg-yw4l">11:09:00</td>
<td class="tg-yw4l">(aa,2)</td>
<td class="tg-yw4l"></td>
<td class="tg-yw4l"></td>
<td class="tg-yw4l">2</td>
</tr>
<tr>
<td class="tg-9hbo">11-12</td>
<td class="tg-yw4l">12:00:00</td>
<td class="tg-yw4l"></td>
<td class="tg-yw4l">aa</td>
<td class="tg-yw4l">1</td>
<td class="tg-yw4l"></td>
</tr>
<tr>
<td class="tg-9hbo" colspan="6">...</td>
</tr>
</table>
Implementation option
---------------------
Considering that the behavior depends on over window behavior, the
implementation will be done by reusing the existing implementation of the over
window functions - done based on processFunction. As mentioned in the
description section, there are 2 options to consider:
1) Using distinct within the aggregates implementation by extending with
distinct aggregates implementation the current aggregates in Flink. For this we
define additional JIRA issues for each implementation to support the distinct
keyword.
2) Using distinct for selection within the process logic when calling the
aggregates. This requires a new implementation of the process Function used for
computing the aggregates. The processFunction will also carry the logic of
taking each element once. For this 2 options are possible. Option 1 (To be
used within the ProcessFunction) trades memory – and would require to create a
hashmap (e.g. mapstate) with binary values to mark if the event was saw
before. This will be created once per window and will be reused across multiple
distinct aggregates. Option 2 trades computation and would require to sort the
window contents and in case of identical elements to eliminate them. The
sorting can be done based on hash values in case the events are non numeric or
composite or do not possess an id to mark the uniqueness. Option 2 is not
prefered for incremental aggregates and should be consider only if certain
aggregates would require a window implementation that recomputes everything
from scratch.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
