suneet-s commented on a change in pull request #10935:
URL: https://github.com/apache/druid/pull/10935#discussion_r592067346



##########
File path: docs/ingestion/compaction.md
##########
@@ -0,0 +1,210 @@
+---
+id: compaction
+title: "Compaction"
+description: "Defines compaction and automatic compaction (auto-compaction or 
autocompaction) as a strategy for segment optimization. Use cases for 
compaction. Describes compaction task configuration."
+---
+
+<!--
+  ~ Licensed to the Apache Software Foundation (ASF) under one
+  ~ or more contributor license agreements.  See the NOTICE file
+  ~ distributed with this work for additional information
+  ~ regarding copyright ownership.  The ASF licenses this file
+  ~ to you under the Apache License, Version 2.0 (the
+  ~ "License"); you may not use this file except in compliance
+  ~ with the License.  You may obtain a copy of the License at
+  ~
+  ~   http://www.apache.org/licenses/LICENSE-2.0
+  ~
+  ~ Unless required by applicable law or agreed to in writing,
+  ~ software distributed under the License is distributed on an
+  ~ "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+  ~ KIND, either express or implied.  See the License for the
+  ~ specific language governing permissions and limitations
+  ~ under the License.
+  -->
+Query performance in Apache Druid depends on optimally sized segments. 
Compaction is one strategy you can use to optimize segment size for your Druid 
database. Compaction tasks read an existing set of segments for a given time 
interval and combine the data into a new "compacted" set of segments. The 
compacted segments are generally larger, but there are fewer of them. Here 
compaction increases performance because fewer segments require less the 
per-segment processing and the memory overhead for ingestion and for querying 
paths.
+
+As a general strategy, compaction is effective when you have data arriving out 
of chronological order resulting in lots of small segments. This often happens, 
for example, if you are appending data using `appendToExisting` for [native 
batch](./native_batch.md) ingestion. Conversely, if you are rewriting your data 
with each ingestion task, you don't need to use compaction.
+
+In some cases you can use compaction to reduce segment size. For example, if a 
misconfigured ingestion task creates oversized segments, you can create a 
compaction task to split the segment files into smaller, more optimally sized 
ones.
+
+See [Segment optimization](../operations/segment-optimization.md) for guidance 
to determine if compaction will help in your environment.
+
+
+## Types of segment compaction
+You can configure the Druid Coordinator to perform automatic compaction, also 
called auto-compaction, for a datasource. Using a segment search policy, the 
coordinator periodically identifies segments for compaction starting with the 
newest to oldest. When segments can benefit from compaction, the coordinator 
automatically submits a compaction task. 
+
+Automatic compaction works in most use cases and should be your first option. 
To learn more about automatic compaction, see [Compacting 
Segments](../design/coordinator.md#compacting-segments).
+
+In cases where you require more control over compaction, you can manually 
submit compaction tasks. For example:
+- Automatic compaction is too slow.
+- You want to force compaction for a specific time range.
+- You want to compact data out of chronological order.
+
+See [Setting up a manual compaction task](#setting-up-manual-compaction) for 
more about manual compaction tasks.
+
+
+## Data handling with compaction
+During compaction, Druid overwrites the original set of segments with the 
compacted set without modifying the data. During compaction Druid locks the 
segments for the time interval being compacted to ensure data consistency.
+
+If an ingestion task needs to write data to a segment for a time interval 
locked for compaction, the ingestion task supersedes the compaction task and 
the compaction task fails without finishing. For manual compaction tasks you 
can adjust the input spec interval to avoid conflicts between ingestion and 
compaction. For automatic compaction, you can set the `skipOffsetFromLatest` 
key to adjustment the auto compaction starting point from the current time to 
reduce the chance of conflicts between ingestion and compaction. See 
[Compaction dynamic 
configuration](../configuration/index.md#compaction-dynamic-configuration) for 
more information.
+
+### Segment granularity handling
+
+Unless you modify the segment granularity in the [granularity 
spec](#compaction-granularity-spec), Druid attempts to retain the granularity 
for the compacted segments. When segments have different segment granularities 
with no overlap in interval Druid creates a separate compaction task for each 
to retain the segment granularity in the compacted segment. If segments have 
different segment granularities before compaction but there is some overlap in 
interval, Druid attempts find start and end of the overlapping interval and 
uses the closest segment granularity level for the compacted segment.
+
+### Query granularity handling
+
+Unless you modify the query granularity in the [granularity 
spec](#compaction-granularity-spec), Druid retains the query granularity for 
the compacted segments. If segments have different query granularities before 
compaction, Druid chooses the finest level of granularity for the resulting 
compacted segment. For example if a compaction task combines two segments, one 
with day query granularity and one with minute query granularity, the resulting 
segment uses minute query granularity.
+
+> In Apache Druid 0.21.0 and prior, Druid sets the granularity for compacted 
segments to the default granularity of `NONE` regardless of the query 
granularity of the original segments.
+
+If you configure query granularity in compaction to go from a finer 
granularity like month to a coarser query granularity like year, then Druid 
overshadows the original segment with coarser granularity. Because the new 
segments have a coarser granularity, running a kill task to remove the 
overshadowed segments for those intervals will cause you to permanently lose 
the finer granularity data.
+
+### Dimension handling
+Apache Druid supports schema changes. Therefore, dimensions can be different 
across segments even if they are a part of the same data source. See [Different 
schemas among 
segments](../design/segments.md#different-schemas-among-segments). If the input 
segments have different dimensions, the resulting compacted segment includes 
all dimensions of the input segments. 
+
+Even when the input segments have the same set of dimensions, the dimension 
order or the data type of dimensions can be different. The dimensions of recent 
segments precede that of old segments in terms of data types and the ordering 
because more recent segments are more likely to have the preferred order and 
data types.
+
+If you want to use your own ordering and types, you can specify a custom 
`dimensionsSpec` in the compaction task spec.
+
+### Rollup
+Druid only rolls up the output segment when `rollup` is set for all input 
segments. See [Roll-up](../ingestion/index.md#rollup) for more details.
+You can check that your segments are rolled up or not by using [Segment 
Metadata Queries](../querying/segmentmetadataquery.md#analysistypes).
+

Review comment:
       I think each one of these sections is currently describing the "what". 
Do you have a plan to talk about the "why"?
   
   For example - You should change segmentGranularity in compaction if you 
ingested data and realized data for that time interval is sparse so having a 
larger segmentGranularity will result in better performance.
   
   Similarly for queryGranularity - the why would be that you no longer need 
fine grained resolution in your data. There should be a big warning that tells 
users they can't go from coarser to finer granularity.
   
   For dimension order - you'd want to change it to get better sorting and 
smaller segments.
   
   I don't know what rollup does on it's own...




----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
[email protected]



---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to