Interesting point, so to speak. @Wail, any chance you could post a
Google maps screenshot showing a visualization of the points in this
dataset on the underlying geographic region? (If the dataset is
shareable in that anonymized form?) I would think an R-tree would still
be good for small-region geo queries - possibly shrinking the candidate
object set by a factor of 10,000 - so still useful - and we also do
index-AND-ing now, so we would also combine that shrinkage by other
index-provided shrinkage on any other index-amenable predicates. I
think the queries are still spatial in nature, and the only AsterixDB
choices for that are R-tree. (We did experiments with things like
Hilbert B-trees, but the results led to the conclusion that the code
base only needs R-trees for spatial data for the forseeable future -
they just work too well and in a no-tuning-required fashion.... :-))
On 9/14/16 12:49 PM, Ahmed Eldawy wrote:
Looks like an interesting case. Just a small question. Are you sure a
spatial index is the right one to use here? The spatial attribute looks
more like a categorization and a hash or B-tree index could be more
suitable. As far as I know, the spatial index in AsterixDB is a secondary
R-tree index which, like any other secondary index, is only good for
retrieving a small number of records. For this dataset, it seems that any
small range would still return a huge number of records.
It is still interesting to further investigate and fix the sort issue but I
mentioned the usage issue for a different perspective.
Thanks
Ahmed
On Wed, Sep 14, 2016 at 10:30 AM Mike Carey <[email protected]> wrote:
☺!
On Sep 14, 2016 1:11 AM, "Wail Alkowaileet" <[email protected]> wrote:
To be exact
I have 2,255,091,590 records and 10,391 points :-)
On Wed, Sep 14, 2016 at 10:46 AM, Mike Carey <[email protected]> wrote:
Thx! I knew I'd meant to "activate" the thought somehow, but couldn't
remember having done it for sure. Oops! Scattered from VLDB, I
guess...!
On 9/13/16 9:58 PM, Taewoo Kim wrote:
@Mike: You filed an issue -
https://issues.apache.org/jira/browse/ASTERIXDB-1639. :-)
Best,
Taewoo
On Tue, Sep 13, 2016 at 9:28 PM, Mike Carey <[email protected]>
wrote:
I can't remember (slight jetlag? :-)) if I shared back to this list
one
theory that came up in India when Wail and I talked F2F - his data
has
a
lot of duplicate points, so maybe something goes awry in that case.
I
wonder if we've sufficiently tested that case? (E.g., what if there
are
gazillions of records originating from a small handful of points?)
On 8/26/16 9:55 AM, Taewoo Kim wrote:
Based on a rough calculation, per partition, each point field takes
3.6GB
(16 bytes * 2887453794 records / 12 partition). To sort 3.6GB, we
are
generating 625 files (96MB or 128MB each) = 157GB. Since Wail
mentioned
that there was no issue when creating a B+ tree index, we need to
check
what SORT process is required by R-Tree index.
Best,
Taewoo
On Fri, Aug 26, 2016 at 7:52 AM, Jianfeng Jia <
[email protected]
wrote:
If all of the file names start with “ExternalSortRunGenerator”, then
they
are the first round files which can not be GCed.
Could you provide the query plan as well?
On Aug 24, 2016, at 10:02 PM, Wail Alkowaileet <[email protected]
wrote:
Hi Ian and Pouria,
The name of the files along with the sizes (there were 625 one of
those
before crashing):
size name
96MB ExternalSortRunGenerator8917133039835449370.waf
128MB ExternalSortRunGenerator8948724728025392343.waf
no files were generated beyond runs.
compiler.sortmemory = 64MB
Here is the full logs
<https://www.dropbox.com/s/k2qbo3wybc8mnnk/log_Thu_Aug_
25_07%3A34%3A52_AST_2016.zip?dl=0>
On Tue, Aug 23, 2016 at 9:29 PM, Pouria Pirzadeh <
[email protected]>
wrote:
We previously had issues with huge spilled sort temp files when
creating
inverted index for fuzzy queries, but NOT R-Trees.
I also recall that Yingyi fixed the issue of delaying clean-up
for
intermediate temp files until the end of the query execution.
If you can share names of a couple of temp files (and their sizes
along
with the sort memory setting you have in
asterix-configuration.xml)
we
may
be able to have a better guess as if the sort is really going
into a
two-level merge or not.
Pouria
On Tue, Aug 23, 2016 at 11:09 AM, Ian Maxon <[email protected]>
wrote:
I think that execption ("No space left on device") is just casted
from
the
native IOException. Therefore I would be inclined to believe it's
genuinely
out of space. I suppose the question is why the external sort is
so
huge.
What is the query plan? Maybe that will shed light on a possible
cause.
On Tue, Aug 23, 2016 at 9:59 AM, Wail Alkowaileet <
[email protected]
wrote:
I was monitoring Inodes ... it didn't go beyond 1%.
On Tue, Aug 23, 2016 at 7:58 PM, Wail Alkowaileet <
[email protected]
wrote:
Hi Chris and Mike,
Actually I was monitoring it to see what's going on:
- The size of each partition is about 40GB (80GB in total
per
iodevice).
- The runs took 157GB per iodevice (about 2x of the
dataset
size).
Each run takes either of 128MB or 96MB of storage.
- At a certain time, there were 522 runs.
I even tried to create a BTree Index to see if that happens as
well.
I
created two BTree indexes one for the *location* and one for the
*caller
*and
they were created successfully. The sizes of the runs didn't
take
anyway
near that.
Logs are attached.
On Tue, Aug 23, 2016 at 7:19 PM, Mike Carey <
[email protected]>
wrote:
I think we might have "file GC issues" - I vaguely remember that
we
don't
(or at least didn't once upon a time) proactively remove
unnecessary
run
files - removing all of them at end-of-job instead of at the
end
of
the
execution phase that uses their contents. We may also have an
"Amdahl
problem" right now with our sort since we serialize phase two
of
parallel
sorts - though this is not a query, it's index build, so that
shouldn't
be
it. It would be interesting to put a df/sleep script on each
of
the
nodes
when this is happening - actually a script that monitors the
temp
file
directory - and watch the lifecycle happen and the sizes
change....
On 8/23/16 2:06 AM, Chris Hillery wrote:
When you get the "disk full" warning, do a quick "df -i" on
the
device
-
possibly you've run out of inodes even if the space isn't all
used
up.
It's
unlikely because I don't think AsterixDB creates a bunch of
small
files,
but worth checking.
If that's not it, then can you share the full exception and
stack
trace?
Ceej
aka Chris Hillery
On Tue, Aug 23, 2016 at 1:59 AM, Wail Alkowaileet <
[email protected]>
wrote:
I just cleared the hard drives to get 80% free space. I still
get
the
same
issue.
The data contains:
1- 2887453794 records.
2- Schema:
create type CDRType as {
id:uuid,
'date':string,
'time':string,
'duration':int64,
'caller':int64,
'callee':int64,
location:point?
}
On Tue, Aug 23, 2016 at 9:06 AM, Wail Alkowaileet <
[email protected]
wrote:
Dears,
I have a dataset of size 290GB loaded in a 3 NCs each of
which
has
2x500GB
SSD.
Each of NC has two IODevices (partitions) in each hard
drive
(i.e
the
total is 4 iodevices per NC). After loading the data, each
Asterix
partition occupied 31GB.
The cluster has about 50% free space in each hard drive
(approximately
about 250GB free space in each hard drive). However, when I
tried
to
create
an index of type RTree, I got an exception that no space left
in
the
hard
drive during the External Sort phase.
Is that normal ?
--
*Regards,*
Wail Alkowaileet
--
*Regards,*
Wail Alkowaileet
--
*Regards,*
Wail Alkowaileet
--
*Regards,*
Wail Alkowaileet
--
*Regards,*
Wail Alkowaileet
Best,
Jianfeng Jia
PhD Candidate of Computer Science
University of California, Irvine
--
*Regards,*
Wail Alkowaileet
