Hi Gaspar,
(this is Jena 4.6.1?)
Not something I recall seeing before.
Are new Data-000_ directories being created?
What's in the log file about backups?
Backups are serialized to one at a time per dataset.
Andy
On 22/11/2022 14:18, Bartalus Gáspár wrote:
Hi Andy,
We’ve just started to run the compaction on our database, but we are
encountering that compaction doesn’t always complete.
Any ideas what could cause this behaviour?
We’re executing the http POST request /$/compact/database_name?deleteOld=true
and subsequently checking the tasks with /$/tasks.
Here is a sample output for the latter, showing that the first 3 attempts
succeeded in ~6 seconds, but the attempts 4, 5 and 6 are not completing.
[ {
"task" : "Compact" ,
"taskId" : "4" ,
"started" : "2022-11-22T14:02:55.216+00:00"
} ,
{
"task" : "Compact" ,
"taskId" : "5" ,
"started" : "2022-11-22T14:05:21.628+00:00"
} ,
{
"task" : "Compact" ,
"taskId" : "6" ,
"started" : "2022-11-22T14:08:17.003+00:00"
} ,
{
"task" : "Compact" ,
"taskId" : "1" ,
"started" : "2022-11-22T13:51:53.060+00:00" ,
"finished" : "2022-11-22T13:51:59.105+00:00" ,
"success" : true
} ,
{
"task" : "Compact" ,
"taskId" : "2" ,
"started" : "2022-11-22T13:54:52.477+00:00" ,
"finished" : "2022-11-22T13:54:58.241+00:00" ,
"success" : true
} ,
{
"task" : "Compact" ,
"taskId" : "3" ,
"started" : "2022-11-22T14:01:33.245+00:00" ,
"finished" : "2022-11-22T14:01:40.070+00:00" ,
"success" : true
}
]
Thanks in advance,
Gaspar
On 14 Jul 2022, at 19:36, Andy Seaborne <a...@apache.org> wrote:
On 07/07/2022 16:19, Lorenz Buehmann wrote:
I think we should wait for Andy here with further input as he's the persons who
basically designed and implemented all the fancy stuff and knows better advice
for sure.
@Andy Did you read the whole discussion and can you verify that it's expected
behavior that lot's of daily updates lead to such a big growth of the node
table files?
Sorry for the delay.
There is no Lucene index by default.
SPO.dat is not nodes table related - it is the base level of the SPO B+Tree.
SPO.idn is the tree above the base level and SPO.bpt keeps the pointers to the
root block and some size information.
The issue looks to be the large numbers of small updates. TDB2 used a
copy-by-write MVCC scheme which means transactions can proceed without needing
latches (database locks) but has the consequence of needing compaction.
TDB1 with Fuseki is worth a try. It does not use the scheme. It does grow but
much more slowly. It is limited in the size of updates it can handle but the
limit is no where near what you describe.
Also worth trying is compaction and deletion
/$/compact/db_name?deleteOld=true
which will delete the old database after compaction (only the one just
compacted. Old ones can be manually deleted).
Andy
On 07.07.22 10:53, Bartalus Gáspár wrote:
Hi Lorenz,
Would you recommend using tdb1 instead of tdb2 for our use case? What would be
the differences?
We are using fuseki 4.5.0 btw.
Gaspar
On 6 Jul 2022, at 14:39, Bartalus Gáspár
<bartalus.gas...@codespring.ro.INVALID> wrote:
Hi,
Most of the updates are DELETE/INSERT queries, i.e
DELETE {?s ?p ?oldValue}
INSERT {?s ?p ?newValue}
WHERE {
OPTIONAL {?s ?p ?oldValue}
#derive ?newValue from somewhere
}
We also have some separate DELETE queries and INSERT queries.
I’ve tried HTTP POST /$/compact/db_name and as a result the files are getting
back to normal size. However, as far as I can tell the old files are also kept.
This is the folder structure I see:
- databases/db_name/Data-0001 - with the old large files
- databases/db_name/Data-0002 - presumably the result of the compact operation
with normal file sizes.
Is there also some operation (http or cli) that would keep only one (the
latest) data folder, i.e. delete the old files from Data-0001?
Gaspar
On 6 Jul 2022, at 12:52, Lorenz Buehmann <buehm...@informatik.uni-leipzig.de>
wrote:
Ok, interesting
so
we have
- 150k triples, rather small dataset
- loaded into 10MB node table files
- 10 updates every 5s
- which makes up to 24 * 60 * 60 / 5 * 10 ~ 200k updates per day
- and leads to 10GB node table files
Can you share the shape of those update queries?
After doing a "compact" operation, the files are getting back to "normal" size?
On 06.07.22 10:36, Bartalus Gáspár wrote:
Hi Lorenz,
Thanks for quick feedback and clarification on lucene indexes.
Here are my answers to your questions:
- We are uploading 7 ttl files to our dataset, where 1 is larger 6Mb, the
others are below 200Kb.
- The overall number of triples after data upload is ~150000.
- We have around 10 SPARQL UPDATE queries that are executed on a regular and
frequent basis, i.e. every 5 seconds. We also have 5 such queries that are
executed each minute. But most of the time they do not produce any outcome,
i.e. the dataset is not altered, and when they do, there are just a couple of
triples that are added to the dataset.
- These *.dat files start from ~10Mb in size, and after a day or so some of
them grow to ~10Gb.
We have ~300 blank nodes, and ~half of the triples have a literal in the object
position, so ~75000.
Best regards,
Gaspar
On 6 Jul 2022, at 10:55, Lorenz Buehmann <buehm...@informatik.uni-leipzig.de>
wrote:
Hi and welcome Gaspar.
Those files do contain the node tables.
A Lucene index is never computed by default and would be contained in Lucene
specific index files.
Can you give some details about the
- size of the files
- the number of triples
- the number triples added/removed/changed
- the frequency of updates
- how much the files grow
- what kind of data you insert? Lots of blank nodes? Or literals?
Also, did you try a compact operation during time?
Lorenz
On 06.07.22 09:40, Bartalus Gáspár wrote:
Hi Jena support team,
We are experiencing an issue with Jena Fuseki databases. In the databases
folder we see some files called SPO.dat, OSP.dat, etc., and the size of these
files are growing quickly. From our understanding these files are containing
the Lucene indexes. We would have two questions:
1. Why are these files growing rapidly, although the underlying data (triples)
are not being changed, or only slightly changed?
2. Can we disable indexing easily, since we are not using full text searches in
our SPARQL queries?
Our usage of Jena Fuseki:
* Start the server with `fuseki-server —port 3030`
* Create databases with HTTP POST to
`/$/datasets?state=active&dbType=tdb2&dbName=db_name`
* Upload ttl files with HTTP POST to /db_name/data
Thanks in advance for your feedback, and if you’d require more input from our
side, please let me know.
Best regards,
Gaspar Bartalus
