I like Mike McCandless's suggestion of controlling which DWPT (and thus segment) an incoming document goes to. I've thought of this before for a different use case grouping documents into segments by the underlying "type" of the document. This could make sense for a use-case that queries by document type, and you don't want to create an index per document type (maybe because the index is too small to warrant it). It could even be used in a kind of soft / hint kind of way -- not an absolute strict separation. For example, say if some subset of DWPTs are known to hold docs of a given type, then add incoming docs of that type to any of those and not the others. But if none exist then just add to any DWPT. I also thought of this sort of thing at the MergePolicy level, but at that point, any mixing of doc types has already occurred and MP can't separate them, it can only combine, though it can try to reduce introducing too much mixing. It would be nice if it were possible to atomically merge some documents in a segment but not the whole segment, thus still leaving the segment in place but with the extracted documents marked deleted. This is similar to "shard splitting" (index splitting) but to do so atomically/transactionally.
~ David Smiley Apache Lucene/Solr Search Developer http://www.linkedin.com/in/davidwsmiley On Sun, Dec 20, 2020 at 10:24 AM Michael McCandless < [email protected]> wrote: > I think the addIndexes approach could work as Haoyu describes! One > IndexWriter per segment in the original source index, using > FilterIndexReader to ... mark all documents NOT in the target segment as > deleted? > > For the final step, you could use addIndexes(Directory[]) which more of > less does a simple file copy of the incoming segment's files. > > But this is a whole extra and costly sounding step, that might undo the > wall clock speedup from the concurrent indexing in the first pass. Maybe > it is still faster net/net than what luceneutil benchmarks, which is > single-threaded-everything (single indexing thread, SerialMergeScheduler, > LogDocMergePolicy)? > > The first option Haoyu listed sounds interesting too! Could we somehow > build a new index, concurrently, but force certain docs to go to certain > in-memory segments (DWPT)? Today the routing of incoming indexing thread > to DWPT is sort of random, but there is indeed a dedicated internal class > that decides that: DocumentsWriterPerThreadPool. And, here is a fun PR > that Adrien is working on to improve how threads are scheduled onto > in-memory segments, to try to create larger initially flushed segments and > less merge pressure as a result: > https://github.com/apache/lucene-solr/pull/1912 > > If we could carefully guide threads to the right DWPT during indexing the > 2nd time, and then use a custom MergePolicy that is also careful to only > merge segments that "belong" together, and the index is sorted, I think you > would get the same segment geometry in the end, and exact same documents in > each segments? This'd likely be nearly as fast as freely building an index > concurrently! It'd be a nice addition to luceneutil benchmarks too, since > now it takes crazy long to build the deterministic index. > > Mike McCandless > > http://blog.mikemccandless.com > > > On Sat, Dec 19, 2020 at 2:50 PM Haoyu Zhai <[email protected]> wrote: > >> Hi Adrien >> I think Mike's comment is correct, we already have index sorted but we >> want to reconstruct a index with exact same number of segments and each >> segment contains exact same documents. >> >> Mike >> AddIndexes could take CodecReader as input [1], which allows us to pass >> in a customized FilteredIndexReader I think? Then it knows which docs to >> take. And then suppose original index has N segments, we could open N >> IndexWriter concurrently and rebuilt those N segments, and at last somehow >> merge them back to a whole index. (I am not quite sure about whether we >> could achieve the last step easily, but that sounds not so hard?) >> >> [1] >> https://lucene.apache.org/core/7_4_0/core/org/apache/lucene/index/IndexWriter.html#addIndexes-org.apache.lucene.index.CodecReader...- >> >> Michael Sokolov <[email protected]> 于2020年12月19日周六 上午9:13写道： >> >>> I don't know about addIndexes. Does that let you say which document goes >>> where somehow? Wouldn't you have to select a subset of documents from each >>> originally indexed segment? >>> >>> On Sat, Dec 19, 2020, 12:11 PM Michael Sokolov <[email protected]> >>> wrote: >>> >>>> I think the idea is to exert control over the distribution of documents >>>> among the segments, in a deterministic reproducible way. >>>> >>>> On Sat, Dec 19, 2020, 11:39 AM Adrien Grand <[email protected]> wrote: >>>> >>>>> Have you considered leveraging Lucene's built-in index sorting? It >>>>> supports concurrent indexing and is quite fast. >>>>> >>>>> On Fri, Dec 18, 2020 at 7:26 PM Haoyu Zhai <[email protected]> wrote: >>>>> >>>>>> Hi >>>>>> Our team is seeking a way of construct (or rebuild) a deterministic >>>>>> sorted index concurrently (I know lucene could achieve that in a >>>>>> sequential >>>>>> manner but that might be too slow for us sometimes) >>>>>> Currently we have roughly 2 ideas, all assuming there's a pre-built >>>>>> index and have dumped a doc-segment map so that IndexWriter would be able >>>>>> to be aware of which doc belong to which segment: >>>>>> 1. First build index in the normal way (concurrently), after the >>>>>> index is built, using "addIndexes" functionality to merge documents into >>>>>> the correct segment. >>>>>> 2. By controlling FlushPolicy and other related classes, make sure >>>>>> each segment created (before merge) has only the documents that belong to >>>>>> one of the segments in the pre-built index. And create a dedicated >>>>>> MergePolicy to only merge segments belonging to one pre-built segment. >>>>>> >>>>>> Basically we think first one is easier to implement and second one is >>>>>> faster. Want to seek some ideas & suggestions & feedback here. >>>>>> >>>>>> Thanks >>>>>> Patrick Zhai >>>>>> >>>>> >>>>> >>>>> -- >>>>> Adrien >>>>> >>>>
