[hibernate-dev] Hibernate Search: massive batch indexing

[Sanne Grinovero]

Hello list,

I've finally finished some performance test about stuff I wanted to double-check
before writing stupid ideas to this list, so I feel I can at last propose
some code to (re)building the index for Hibernate Search.

The present API of Hibernate Search provides a nice and safe
transactional "index(entity)",
but even when trying several optimizations it doesn't reach the speed
of an unsafe (out of transaction) indexer we use in our current
production environment.
Also reading the forum it appears that much people are having
difficulties in using
the current API, even with a good example in the reference documentation
some difficulties arise with Seam's transactions and with huge data sets.
(I'm NOT saying something is broken, just that you need a lot of expertise
to get it going)

SCENARIO
=======

* Developers change an entity and want to test the effect on the index
structure,
  thay want do to search experiments with the new fields.
* A production system is up(down)graded to a new(old) release,
involving index changes.
  (the system is "down for maintance" but the speed is crucial)
* Existing index is corrupted/lost. (Again, speed to recover is critical)
* A Database backup is restored, or data is changed by other jobs.
* Some crazy developer like me prefers to disable H.Search's event
listeners for some reason.
  (I wouldn't generally recommend it, but have met other people who
have a reasonable
  argument to do this. Also in our case it is a feature as new entered
books will be
  available for loans only from the next day :D)
* A Lucene update breaks the index format (not so irrationale as they
just did on trunk).

PERFORMANCE
=======

In simple use cases, such as less than 1000 entities and not too much
relationships,
the exising API outperforms my prototype, as I have some costly setup.
In more massive tests the setup costs are easily recovered by a much
faster indexing speed;
I have many data I could send, I'll just show some and keep the details simple:

entity "Operator": standard complexity, involves loading of +4 objs, 7
field affect index
entity "User": moderate complexity, involves loading of +- 20 objs, 12
affect index data
entity "Modern": high complexity, loading of 44 entities, many are
"manyToMany", 25 affect index data

On my laptop (dual core, local MySQL db):
type		Operator		User		Modern
number		560			100.000		100.000
time-current	0,23 secs		45''		270.3''
time-new	0,43 secs		30''		190''

On a staging server (4 core Xeon with lots of ram and dedicated DB server):
type		Operator		User		Modern
number		560			200.000		4.000.000
time-current	0,09 secs		130''		5h20'
time-new	0,25 secs		22''		19'

[benchmark disclaimer:
These timings are meant to be relative to each other for my particular
code version, I'm not an expert of Java benchmarking at all.
Also unfortunately I can't really access the same hardware for each tests.
I used all possible tweaks I am aware of in Hibernate Search, actually
enabling new needed params to make the test as fair as possible.]

Examining the numbers:
   with current recommended H.Search strategy I can index 560 simple entities
in 0,23 seconds; quite fast and newbe users will be impressed.
   At the other extreme, we index 4 million complex items, but I need more
than 5 hours to do that; this is more like real use case and it could
scare several developers.
   Unfortunately I don't have a complete copy of the DB on my laptop,
but looking at the numbers it looks like my laptop could finish
in 3 hours, nearly double the speed of our more-than-twice fast server.
(yes I've had several memory leaks :-) but they're solved now)
   The real advantage is the round-trip to database: without multiple
threading each lazy loaded collection somehow annotated to be indexed
massively slows down the whole process; If you look at both DB an AS
servers, they have very low resource usage confirming this, while my laptop
stays at 70% cpu (and killing my harddrive) because he has data available
locally, producing a constant feed of strings to my index.
   When using the new prototype (about 20 threads in 4 different pools)
I get the 5hours down to less than 20minutes; Also I can start the
indexing of all 7 indexable types in parallel and it will stay around 20minutes.
The "User" entity is not as complex as Modern (less lazy loaded data)
but confirms the same numbers.

ISSUES
=======
About the current version I've ready.
It is not a complete substitute of the current one and is far from perfect;
currently these limitations apply but could be easily solved:
(others I am not aware of not listed :-)

A) I need to "read" some hints for each entity; I tinkered with a new
annotation,
   configuration properties should work but are likely to be quite
verbose (HQL);
   basically I need some hints about fetch strategies appopriate
   for batch indexing, which are often different than normal use cases.

B) Hibernate Search's indexing of related entities was not available
when I designed it.
   I think this change will probably not affect my code,  but I still need to
   verify the functionality of IndexEmbedded.

C) It is finetuned for our entities and DB, many variables are configurable but
   some stuff should be made more flexible.

D) Also index sharding didn't exist at the time, I'll need to change some stuff
   to send the entities to the correct index and acquire the appropriate locks.

The next limitations is not easy to solve, I have some ideas but no one I liked.

E) It is not completely safe to use it during other data modification;
It's not a problem in our
   current production but needs much warning in case other people
wants to use it.
   The best solution I could think of is to lock the current workqueue
of H.Search,
   so to block execution of work objects in the queue and resume the
execution of
   this work objects after batch indexing is complete.
   If some entity disappears (removed from DB but a reference is in
the queue) it
   can easily be skipped, if I index "old version" of some other data it will be
   fixed when scheduled updates from H.S. eventlisteners are resumed;
   (and the same for new entities).
   It would be nice to share the same database transaction during the
whole process,
   but as I use several threads and many separate sessions I think
this is not possible
   (this is the best place to ask I think;-)

GOING PRACTICAL
===============
if (cheater) goto :top

A nice evictAll(class) exists, I would like to add indexAll(class).
It would be nice to provide non-blocking versions, maybe overloading:
indexAll(Class clazz, boolean block)
or provide a Future as return object, so people could wait for one
or more indexAll requests if they want to.
There are many parameters to tweak the indexing process, so I'm
not sure if we should put them in the properties, or have a parameters-
wrapper object indexAll(Class class, Properties prop), or
something like makeIndexer(Class class) returning a complex object
with several setters for finetuning and start() and awaitTermination()
methods.

the easy part
--------------
This part is easy to do as I have it working well, it is a pattern
involving several executors; the size of each threadPool and of the
linking queues between them gives the good balance to achieve the
high throughput.
First the entities are counted and divided in blocks, these ranges are fed to
N scrollables opened in N threads, each thread begins iterating on the
list of entities and feeds detached entities to the next Pool using
BlockingQueues.
In the next pool the entities are re-attached using Lock.none, readonly, etc..
(and many others you may want to tell me) and we get and appropriate
DocumentBuilder from the SearchFactory to transform it into a Lucene Document;
this pool is usually the slowest as it has to initialize many lazy fields,
so there are more threads here.
Produced documents go to a smaller pool (best I found was for 2-3 threads)
were data is concurrently written to the IndexWriter.
There's an additional thread for resource monitoring to produce some hints
about queue sizing and idle threads, to do some finetune and to see instant
speed reports in logs when enabled.
For shutdown I use the "poison pill" pattern, and I usually get rid of all
threads and executors when I'm finished.
It needs some adaption to take into account of latest Search features
such as similarity, but is mostly beta-ready.

the difficult part
-------------------
Integrating it with the current locking scheme is not really difficult,
also because the goal is to minimize downtime so I think some downtime
should be acceptable.
It would be very nice however integrate this pattern as the default
writer for indexes, even "in transaction"; I think it could be possible
even in synchronous mode to split the work of a single transaction across
the executors and wait for all the work be done at commit.
You probably don't want to see the "lots of threads" meant for batch indexing,
but the pools scale quite well to adapt themselves to the load,
and it's easy (as in clean and maintainable code) to enforce resource limits.
When integrating at this level the system wouldn't need to stop regular
Search activity.

any questions? If someone wanted to reproduce my benchmarks I'll
be glad to send my current code and DB.

kind regards,
Sanne