OK, back to ERIndexing discussion for my 5-minute "break" ......
Anjo, does it matter that p2 updates the index before p1 since updating the index consists of reading the EO from the database?
In the example shown below "p2 updates index" would update the index to reflect the EO after "p2 changes data" and the out of FIFO order "p1 updates index" would simply cause the index to again reflect the EO after "p2 changes data". In other words, because the of the latency or lack of synchronization, the index never reflected the changes of "p1 changes data", but who cares since a few milliseconds later p1's changes were replaced in the database by p2's changes anyway.
Corrrect me if I am not thinking logically here, but the example below should be:
p1 changes EO1, p2 changes EO1, p2 updates lucene 'Document' for EO1 from database, p1 updates lucene 'Document' for EO1 from database.
Result, lucene Document for EO1 reflects the last state of the database - all is good?
-Kieran
On Oct 19, 2009, at 9:23 AM, Anjo Krank wrote: It's not a matter of thread safety, it's a matter of data in lucene being the same as in your DB. When you run with multiple instances and have heavy edits, you can easily construct a case like: p1 changes data, p2 changes data, p2 updates index, p1 updates index. When you only have one edit app, that's obviously OK (and I'm using a queue for the single app mode anyway). Cheers, Anjo Am 19.10.2009 um 15:06 schrieb David LeBer:
On 2009-10-19, at 8:26 AM, David LeBer wrote:
On 2009-10-19, at 8:13 AM, Mike Schrag wrote:
it's not like you wouldn't have the exact same problems in lucene-proper, though ...
We are using ERIndexing for a multiple instance single server deployment. However, the app is readonly for the indexed EOs, there is an admin app that writes, but that is only a single instance.
And based on the Lucene docs, it's writers and readers are thread and process safe, which means that multiple writers can access the same index file.
Doug Cutting has posted on the topic of thread safety a couple of times. Indexing and searching are not only thread safe, but process safe. What this means is that:
• Multiple index searchers can read the lucene index files at the same time.
• An index writer or reader can edit the lucene index files while searches are ongoing
• Multiple index writers or readers can try to edit the lucene index files at the same time (it's important for the index writer/reader to be closed so it will release the file lock).
Not sure how well this works in practice and/or how file system dependent it is for the file system locks to function correctly.
On Oct 19, 2009, at 7:50 AM, Gustavo Pizano wrote:
NICE!, now my hopes are gone!.
so I guess I must make use de facto lucene framework. and follow the examples in LIA?
ok.. what can one do... :(
thx
G.
On Mon, Oct 19, 2009 at 1:45 PM, Anjo Krank <email@hidden> wrote:
Be aware that ERIndexing is only an experiment (and was write-only code, I don't use it yet). In particular it has several severe drawbacks:
- it doesn't really handle multiple instances (possibly) or servers (definitely). That means, for the cases where you actually *do* need the speed of lucene, ie. high-traffic, high-volume which means many servers, you can't use it as is. At least the auto-indexing won't work without some central notification point that actually does the indexing and then redistribute the indexes.
If you don't account for that, your indexes won't really match your DB, which means that you will find the wrong stuff super-fast...
I don't have a good solution to this, maybe someone who actually uses it might.
- The DB store for the indexes was an experiment to fix at least the redistribution problem, but this was truly write only, so use at your own risk.
- it duplicates your DB indexes and depending on your DB type and query, your query to resolve the faults probably won't be that much faster than a normal query would have been.
- it should really be an EO adaptor instead, which would mean that you could use it in a simple displayGroup. But then again, one of the main points in Lucene is that you don't really need a strict schema to work with it - although you'll probably have one.
Cheers, Anjo
|