On Sep 22, 2009, at 3:44 PM, Michael McCandless wrote:
On Tue, Sep 22, 2009 at 2:53 PM, Grant Ingersoll
<gsing...@apache.org> wrote:
One of the pieces I still am missing from all of this is why isn't
IW.getReader() now just the preferred way of getting a IndexReader
for all applications other than those that are completely batch
oriented?
Why bother with IndexReader.reopen()?
I agree, most apps should simply use getReader, as long as they're
running in the same JVM as the IndexWriter, and, they are holding the
IW open anyway.
But, the returned reader is read-only, so you can't use it to change
norms, do deletes, etc.
Yeah, but an IW can do deletes, and if the this IR is coupled to it
anyway...
The API really shouldn't be marked expert. I'll go remove that...
Lucene has, in fact, always been about incremental updates (since
there are commercial systems out there that require complete
re-indexing)
True, for writing. But for reading, reopening a reader was very
costly before 2.9 because FieldCache entry had to be fully recomputed.
So, switching to per-segment search/collect in 2.9 was the biggest
step to reducing NRT reopen latency.
and that getting IR.reopen to perform is just a matter of tuning
one's application in regards to reads and writes vs. having to do
all this work in the IndexWriter that now tightly couples the
IndexReader to the IndexWriter.
The integration with IndexWriter allows a reader to access segments
that haven't yet been committed to the index. This saves fsync()'ing
the written files, saves writing a new segments_N file, saves flushing
deletes to disk and then reloading them (we just share the BitVector
directly in RAM now). On many OS/filesystems fsync is surprisingly
costly.
LUCENE-1313, the next step for NRT, further reduces NRT reopen latency
by allowing the small segments to remain in RAM, so when reopening
your NRT reader after smallish add/deletes no IO is incurred.
Beyond LUCENE-1313 we've discussed making IndexWriter's RAM buffer
directly searchable, so you don't pay the cost of pinching a new
segment when an NRT reader is reopened.
Really we only need to further improve the approach here if the
existing performance proves inadequate... in my limited testing the
performance was excellent.
Though, our inability to prioritize IO and control the OS's IO cache,
from java, are likely far bigger impacts on our NRT performance at
this point, than further improvements in our impl. I'd love to see a
Directory impl that "emulates" IO prioritization by making merging IO
wait whenever search IO is live. I think we need a JNI extension that
taps into madvise/posix_fadvise, when possible.
FWIW, I still don't like the coupling of the two. I think it would
be better if IW allowed you to get a Directory (or some other
appropriate representation) representing the in memory segment that
can then easily be added to an existing Searcher/Reader. This would
at least decouple the two and instead use the common data structure
they both already share, i.e. the Directory. Whether this is doable
or not, I am not sure.
I agree the coupling is overkill.
But Directory is too low... we could probably get by with a class that
holds the SegmentReader cache (currently IndexWriter.ReaderPool), and
the "current" segmentInfos. IW would interact with this class to get
the readers it needs, for applying deletes, merging, as well as
posting newly flushed but not yet committed segments, and IR would
then pull from this class to get the latest segments in the index and
to checkout the readers.
Not sure why Directory, a public well-known class, is considered too
low (I thought you would say too high!) versus inner classes that
assume an IndexWriter. The reason I chose Directory is because it is
the common thing already shared between the two and it is already a
public, well-known class that requires no extra understanding by
users. It's a first class citizen. By reusing it, apps can be
agnostic about where it came from, versus having to wire in all this
new stuff to handle ReaderPools, etc. versus simply reusing the
directory stuff.
-Grant
---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: java-dev-h...@lucene.apache.org
