Uploaded image for project: 'Lucene - Core'
  1. Lucene - Core
  2. LUCENE-4339

Allow deletions on Lucene3x codecs again

Details

    • Bug
    • Status: Closed
    • Blocker
    • Resolution: Fixed
    • 4.0-BETA
    • 4.0
    • core/index
    • None
    • New

    Description

      On dev@lao Hoss reported that a user in Solr was not able to update or delete documents in his 3.x index with Solr 4:

      On the solr-user list, Dirk Högemann recently mentioned a problem he was seeing when he tried upgrading his existing solr setup from 3.x to 4.0-BETA. Specifically this exception getting logged...

      http://find.searchhub.org/document/cdb30099bfea30c6

      auto commit error...:java.lang.UnsupportedOperationException: this codec can only be used for reading
      at org.apache.lucene.codecs.lucene3x.Lucene3xCodec$1.writeLiveDocs(Lucene3xCodec.java:74)
      at org.apache.lucene.index.ReadersAndLiveDocs.writeLiveDocs(ReadersAndLiveDocs.java:278)
      at org.apache.lucene.index.IndexWriter$ReaderPool.release(IndexWriter.java:435)
      at org.apache.lucene.index.BufferedDeletesStream.applyDeletes(BufferedDeletesStream.java:278)
      at org.apache.lucene.index.IndexWriter.applyAllDeletes(IndexWriter.java:2928)
      at org.apache.lucene.index.IndexWriter.maybeApplyDeletes(IndexWriter.java:2919)
      at org.apache.lucene.index.IndexWriter.prepareCommit(IndexWriter.java:2666)
      at org.apache.lucene.index.IndexWriter.commitInternal(IndexWriter.java:2793)
      at org.apache.lucene.index.IndexWriter.commit(IndexWriter.java:2773)
      at org.apache.solr.update.DirectUpdateHandler2.commit(DirectUpdateHandler2.java:531)
      at org.apache.solr.update.CommitTracker.run(CommitTracker.java:214)

      Dirk was able to work arround this by completely re-indexing, but it seemed strange to me that this would happen.

      My understanding is that even though an IndexUpgrader tool was now available, it wasn't going to be required for users to use it when upgrading from 3.x to 4.x. Explicitly upgrading the index format might be a good idea, and might make hte index more performant, but as I understood it, the way things had been implemented with codecs explicitly upgrading the index format wasn't strictly neccessary, and that users should be able to upgrade their lucene apps same way that was supported with other index format upgrades in the past: the old index can be read, and as changes are made new segments will be re-written in the new format. (Note in
      particular: at the moment we don't mention IndexUpgrader in MIGRATE.txt at all.)

      It appears however, based on this stack trace and some other experiements i tried, that any attempts to "delete" documents in a segment that is using the Lucene3xCodec will fail.

      This seems like a really scary time bomb sitaution, because if you upgrade, things will seem to be working – you can even add documents, and depending on the order that you do things, some "old" segments may get merged and use the new format, so some deletes of "old" documents (in those merged/upgraded) segments may work, but then somewhere down the road, you may try to a delete that affects docs in a still un-merge/upgraded segment, and that delete will fail – 5 minutes later, if another merge has happened, attempting to do the exact same delete may succeed.

      All of which begs the question: is this a known/intended limitation of the Lucene3xCodec, or an oversight in the Lucene3xCodec?

      if it's expected, then it seems like we should definitely spell out this limitation in MIGRATE.txt and advocate either full rebuilds, or the use of IndexUpgrader for anyone who's indexes are non-static.

      On the Solr side of things, i think we should even want to consider automaticly running IndexUpgrader on startup if we detect that the Lucene3xCodec is in use to simplify things – we can't even suggest running "optimize" as a quick/easy way to force and index format upgrade because if the 3x index as already optimized then it's a no-op and the index stays in the 3x format.

      Robert said, that this is a wanted limitation (in fact its explicitely added to the code, without that UOE it "simply works"), but I disagree here and lots of other people:

      In the early days (I mean in the time when it was already read only until we refactored the IndexReader.delete()/Codec stuff), this was working, because the LiveDocs were always handled in a special way. Making it now 100% read-only is in my opinion very bad, as it does not allow to update documents in a 3.x index anymore, so you have no chance, you must run IndexUpgrader.

      The usual step like opening old Index and adding documents works (because the new documents are added always to new segment), but the much more usual IW.updateDocument() which is commonly used also to add documents fails on old Indexes. This is a no-go, we have to fix this. If we allow the trick with updating LiveDocs on 3.x codec, for the end-user the "read-only" stuff in Lucene3x codec would be completely invisible, as he can do everything IndexWriter provides. The other horrible things like changing norms is no longer possible, so deletes are the only thing that affects here. The read-only ness of Lucene3x codec would only be visible to the user when someone tries to explicitly create an index with Lucene3x codec. And I understood the CHANGES/MIGRATE.txt exactly as that.

      On the list, Robert added a simple patch, reverting the UOE in Lucene3xCodec, so the LiveDocs format is RW again.

      Attachments

        1. LUCENE-4339.patch
          3 kB
          Uwe Schindler
        2. LUCENE-4339.patch
          5 kB
          Michael McCandless

        Activity

          People

            Unassigned Unassigned
            uschindler Uwe Schindler
            Votes:
            0 Vote for this issue
            Watchers:
            2 Start watching this issue

            Dates

              Created:
              Updated:
              Resolved: