[jira] [Comment Edited] (LUCENE-9529) Larger stored fields block sizes mean we're more likely to disable optimized bulk merging

["Robert Muir (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/LUCENE-9529?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17197328#comment-17197328 ] 

Robert Muir edited comment on LUCENE-9529 at 9/17/20, 2:23 AM:
---------------------------------------------------------------

The current code tracks the total number of chunks, and the total number of "dirty" (incomplete) chunks. 

Then we compute "tooDirty" like this:
{code}
  /** 
   * Returns true if we should recompress this reader, even though we could bulk merge compressed data 
   * <p>
   * The last chunk written for a segment is typically incomplete, so without recompressing,
   * in some worst-case situations (e.g. frequent reopen with tiny flushes), over time the 
   * compression ratio can degrade. This is a safety switch.
   */
  boolean tooDirty(CompressingStoredFieldsReader candidate) {
    // more than 1% dirty, or more than hard limit of 1024 dirty chunks
    return candidate.getNumDirtyChunks() > 1024 || 
           candidate.getNumDirtyChunks() * 100 > candidate.getNumChunks();
  }
{code}

Maybe to be more fair, we could use a similar formula but track numDirtyDocs and compare with numDocs (we know this value already)? We could still keep a safety-switch such as 1024 dirty chunks to avoid some worst-case scenario, but just change the ratio at least. 


was (Author: rcmuir):
The current code tracks the total number of chunks, and the total number of "dirty" (incomplete) chunks. 

Then we compute "tooDirty" like this:
{code}
  /** 
   * Returns true if we should recompress this reader, even though we could bulk merge compressed data 
   * <p>
   * The last chunk written for a segment is typically incomplete, so without recompressing,
   * in some worst-case situations (e.g. frequent reopen with tiny flushes), over time the 
   * compression ratio can degrade. This is a safety switch.
   */
  boolean tooDirty(CompressingStoredFieldsReader candidate) {
    // more than 1% dirty, or more than hard limit of 1024 dirty chunks
    return candidate.getNumDirtyChunks() > 1024 || 
           candidate.getNumDirtyChunks() * 100 > candidate.getNumChunks();
  }
{noformat}

Maybe to be more fair, we could use a similar formula but track numDirtyDocs and compare with numDocs (we know this value already)? We could still keep a safety-switch such as 1024 dirty chunks to avoid some worst-case scenario, but just change the ratio at least. 

> Larger stored fields block sizes mean we're more likely to disable optimized bulk merging
> -----------------------------------------------------------------------------------------
>
>                 Key: LUCENE-9529
>                 URL: https://issues.apache.org/jira/browse/LUCENE-9529
>             Project: Lucene - Core
>          Issue Type: Improvement
>            Reporter: Adrien Grand
>            Priority: Minor
>
> Whenever possible when merging stored fields, Lucene tries to copy the compressed data instead of decompressing the source segment to then re-compressing in the destination segment. A problem with this approach is that if some blocks are incomplete (typically the last block of a segment) then it remains incomplete in the destination segment too, and if we do it for too long we end up with a bad compression ratio. So Lucene keeps track of these incomplete blocks, and makes sure to keep a ratio of incomplete blocks below 1%.
> But as we increased the block size, it has become more likely to have a high ratio of incomplete blocks. E.g. if you have a segment with 1MB of stored fields, with 16kB blocks like before, you have 63 complete blocks and 1 incomplete block, or 1.6%. But now with ~512kB blocks, you have one complete block and 1 incomplete block, ie. 50%.
> I'm not sure how to fix it or even whether it should be fixed but wanted to open an issue to track this.



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

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org