Elasticsearch _bulk periodically unavailable through Python API, but seems fine otherwise

Elastic Stack Elasticsearch
1

We're currently working with re-indexing one of our indexes on one of our
test servers, and we appear to be having issues. The operation should be
fairly obvious, and there's not a lot of data -- around 320.000 total
documents in the index, at around 1.15GB of data. We scan from the old
index, and bulk index 1000 documents at a time (we've also tried 500
documents at a time to no avail). Since our production indexes are more
around the 30GB size, we're a bit discouraged by hitting this wall.

Once in a while, we get a random "resource not available" error, but they
usually resolves themselves and go away -- however, after indexing around
260.000 documents, the Python script starts behaving weirdly, throwing a
lot of resource not available errors. A full gist is
here: https://gist.github.com/HenrikOssipoff/ac63060d4c18182719dd (note
that this is Python 3.4, so chained exceptions and everything -- the top
exception is not actually the important one).

Now, thing is. During this time, Elasticsearch itself seems fine, and the
system is generally available for querying. The JVM heap is fine, and the
Elasticsearch log itself has absolutely no warnings or errors (log level
INFO enabled).

Our next course of action will be try to doing this by just cURLs outside
of Python, or even enabling debug logging in Elasticsearch, but we wanted
to see if anyone has some ideas for what could be the course.

It seems logical, that since it only happens every one in a while, and gets
worse the more documents we index, it has something to do with the amount
of _bulk indexes we do, but we can't seem to find anything in logs or stats
that support this.

Any help appreciated, or any pointers in the right direction.

Thank you!

Regards,
Henrik Ossipoff Hansen

--
You received this message because you are subscribed to the Google Groups "elasticsearch" group.
To unsubscribe from this group and stop receiving emails from it, send an email to elasticsearch+unsubscribe@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/elasticsearch/2afa6d99-b9d4-42a2-9f4e-b717a3a41251%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

2

Hello Henrik,

My guess would be that the timeouts you are seeing in python are
causing this - when the python client encounters a timeout, it retries
the request, thinking something went wrong. Thus if your timeout is
too small it can actually lead to spamming the cluster - when the
python client encounters a timeout the request is still running in the
server so each retry will spawn another bulk request being processed.

The solution should be to make the timeout for your python client higher.

I am starting to consider this a bug so I created a ticket to fix it,
you can track it on github:

Hope this helps,
Honza

On Tue, Sep 23, 2014 at 9:48 PM, Henrik Ossipoff Hansen
henrik.ossipoff@gmail.com wrote:

We're currently working with re-indexing one of our indexes on one of our
test servers, and we appear to be having issues. The operation should be
fairly obvious, and there's not a lot of data -- around 320.000 total
documents in the index, at around 1.15GB of data. We scan from the old
index, and bulk index 1000 documents at a time (we've also tried 500
documents at a time to no avail). Since our production indexes are more
around the 30GB size, we're a bit discouraged by hitting this wall.

Once in a while, we get a random "resource not available" error, but they
usually resolves themselves and go away -- however, after indexing around
260.000 documents, the Python script starts behaving weirdly, throwing a lot
of resource not available errors. A full gist is here:
Traceback on bulk indexing · GitHub (note that this
is Python 3.4, so chained exceptions and everything -- the top exception is
not actually the important one).

Now, thing is. During this time, Elasticsearch itself seems fine, and the
system is generally available for querying. The JVM heap is fine, and the
Elasticsearch log itself has absolutely no warnings or errors (log level
INFO enabled).

Our next course of action will be try to doing this by just cURLs outside of
Python, or even enabling debug logging in Elasticsearch, but we wanted to
see if anyone has some ideas for what could be the course.

It seems logical, that since it only happens every one in a while, and gets
worse the more documents we index, it has something to do with the amount of
_bulk indexes we do, but we can't seem to find anything in logs or stats
that support this.

Any help appreciated, or any pointers in the right direction.

Thank you!

Regards,
Henrik Ossipoff Hansen

--
You received this message because you are subscribed to the Google Groups
"elasticsearch" group.
To unsubscribe from this group and stop receiving emails from it, send an
email to elasticsearch+unsubscribe@googlegroups.com.
To view this discussion on the web visit
https://groups.google.com/d/msgid/elasticsearch/2afa6d99-b9d4-42a2-9f4e-b717a3a41251%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--
You received this message because you are subscribed to the Google Groups "elasticsearch" group.
To unsubscribe from this group and stop receiving emails from it, send an email to elasticsearch+unsubscribe@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/elasticsearch/CABfdDioPUGyaJQmOXPv9SUoTfynDbTfnvfAQnrQGGhE_NVr1GA%40mail.gmail.com.
For more options, visit https://groups.google.com/d/optout.

3

Hi Honza,

Hoped you would be lurking around here :slight_smile:

I almost feel ashamed for not thinking about that myself - it makes totally
sense. The default is 10 seconds as I recall, and we haven't tampered with
this in this particular one of our projects. It also makes sense that we
are seeing this type of error more on our test cluster, since it has much
worse hardware than our production nodes.

Moreover, your comment has likely put a point on another issue we see in a
different application, where we have changed the default timeout, but to
something I'm now realising is likely way too low.

Thank you -- I'll be monitoring the issue for sure.

I'll report back once we get our test set up again in a few hours.

Regards,
Henrik

Den tirsdag den 23. september 2014 22.59.32 UTC+2 skrev Honza Král:

Hello Henrik,

My guess would be that the timeouts you are seeing in python are
causing this - when the python client encounters a timeout, it retries
the request, thinking something went wrong. Thus if your timeout is
too small it can actually lead to spamming the cluster - when the
python client encounters a timeout the request is still running in the
server so each retry will spawn another bulk request being processed.

The solution should be to make the timeout for your python client higher.

I am starting to consider this a bug so I created a ticket to fix it,
you can track it on github:

Timeout shouldn't cause a retry · Issue #128 · elastic/elasticsearch-py · GitHub

Hope this helps,
Honza

On Tue, Sep 23, 2014 at 9:48 PM, Henrik Ossipoff Hansen
<henrik....@gmail.com <javascript:>> wrote:

We're currently working with re-indexing one of our indexes on one of
our
test servers, and we appear to be having issues. The operation should be
fairly obvious, and there's not a lot of data -- around 320.000 total
documents in the index, at around 1.15GB of data. We scan from the old
index, and bulk index 1000 documents at a time (we've also tried 500
documents at a time to no avail). Since our production indexes are more
around the 30GB size, we're a bit discouraged by hitting this wall.

Once in a while, we get a random "resource not available" error, but
they
usually resolves themselves and go away -- however, after indexing
around
260.000 documents, the Python script starts behaving weirdly, throwing a
lot
of resource not available errors. A full gist is here:
Traceback on bulk indexing · GitHub (note that
this
is Python 3.4, so chained exceptions and everything -- the top exception
is
not actually the important one).

Now, thing is. During this time, Elasticsearch itself seems fine, and
the
system is generally available for querying. The JVM heap is fine, and
the
Elasticsearch log itself has absolutely no warnings or errors (log level
INFO enabled).

Our next course of action will be try to doing this by just cURLs
outside of
Python, or even enabling debug logging in Elasticsearch, but we wanted
to
see if anyone has some ideas for what could be the course.

It seems logical, that since it only happens every one in a while, and
gets
worse the more documents we index, it has something to do with the
amount of
_bulk indexes we do, but we can't seem to find anything in logs or stats
that support this.

Any help appreciated, or any pointers in the right direction.

Thank you!

Regards,
Henrik Ossipoff Hansen

--
You received this message because you are subscribed to the Google
Groups
"elasticsearch" group.
To unsubscribe from this group and stop receiving emails from it, send
an
email to elasticsearc...@googlegroups.com <javascript:>.
To view this discussion on the web visit

https://groups.google.com/d/msgid/elasticsearch/2afa6d99-b9d4-42a2-9f4e-b717a3a41251%40googlegroups.com.

For more options, visit https://groups.google.com/d/optout.

--
You received this message because you are subscribed to the Google Groups "elasticsearch" group.
To unsubscribe from this group and stop receiving emails from it, send an email to elasticsearch+unsubscribe@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/elasticsearch/4e98937d-9c5c-4ffb-8e90-95796deb6293%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.