Making Ignite In-Memory File System mountable

[dsetrakyan]

I have noticed a Jira filed to make IGFS mountable:
https://issues.apache.org/jira/browse/IGNITE-854

The integration proposed is via Fuse framework which would allow us mount
IGFS just like a regular Linux file system.

This seems like a pretty big undertaking. Even though it sounds pretty cool
in theory, it would be interesting to find out form the community if
everyone else finds it useful and maybe suggest some potential use cases
for it.

Also, seems like Fuse framework is not being actively developed. Last
release came out in July, 2013. How safe is it to use this framework?

D.

[Marko Jevtic]

Sorry for dropping like this in the topic :)

But I think it would be extremely useful if we Ignite would provide an API to virtually expose structures in cache.
Like computing on fly (on access) csv/excel files or pdf reports from cache, besides the regular usage to access Ignite’s file system.

Just an idea.

> On May 5, 2015, at 3:58 AM, Dmitriy Setrakyan <[hidden email]> wrote:
>
> I have noticed a Jira filed to make IGFS mountable:
> https://issues.apache.org/jira/browse/IGNITE-854
>
> The integration proposed is via Fuse framework which would allow us mount
> IGFS just like a regular Linux file system.
>
> This seems like a pretty big undertaking. Even though it sounds pretty cool
> in theory, it would be interesting to find out form the community if
> everyone else finds it useful and maybe suggest some potential use cases
> for it.
>
> Also, seems like Fuse framework is not being actively developed. Last
> release came out in July, 2013. How safe is it to use this framework?
>
> D.

[dsetrakyan]

On Mon, May 4, 2015 at 7:08 PM, Marko Jevtic <[hidden email]> wrote:

> Sorry for dropping like this in the topic :)
>
> But I think it would be extremely useful if we Ignite would provide an API
> to virtually expose structures in cache.
> Like computing on fly (on access) csv/excel files or pdf reports from
> cache, besides the regular usage to access Ignite’s file system.
>

Hm... Are you suggesting parallel processing of different sections of PDF
files from different compute jobs, or have a single compute job process the
whole file?

>
> Just an idea.
>
> > On May 5, 2015, at 3:58 AM, Dmitriy Setrakyan <[hidden email]>
> wrote:
> >
> > I have noticed a Jira filed to make IGFS mountable:
> > https://issues.apache.org/jira/browse/IGNITE-854
> >
> > The integration proposed is via Fuse framework which would allow us mount
> > IGFS just like a regular Linux file system.
> >
> > This seems like a pretty big undertaking. Even though it sounds pretty
> cool
> > in theory, it would be interesting to find out form the community if
> > everyone else finds it useful and maybe suggest some potential use cases
> > for it.
> >
> > Also, seems like Fuse framework is not being actively developed. Last
> > release came out in July, 2013. How safe is it to use this framework?
> >
> > D.
>
>

[Marko Jevtic]

No, but that’s a good idea. Not sure if PDF format can be streamed/processed parallely.

I thought, if you would bind FUSE to expose Ignite’s file system, I would leave an API for the developer, to let him expose the data in Ignite. Since it’s a user-space file system, we can expose directories/files as we want.
I see that as useful, as one can later mount the file system via smb/cifs to other platforms and have access to in-memory real time system with reports etc… Something similar that Alfresco is doing with JLan.

I might be wrong :)

> On May 5, 2015, at 4:12 AM, Dmitriy Setrakyan <[hidden email]> wrote:
>
> On Mon, May 4, 2015 at 7:08 PM, Marko Jevtic <[hidden email]> wrote:
>
>> Sorry for dropping like this in the topic :)
>>
>> But I think it would be extremely useful if we Ignite would provide an API
>> to virtually expose structures in cache.
>> Like computing on fly (on access) csv/excel files or pdf reports from
>> cache, besides the regular usage to access Ignite’s file system.
>>
>
> Hm... Are you suggesting parallel processing of different sections of PDF
> files from different compute jobs, or have a single compute job process the
> whole file?
>
>
>>
>> Just an idea.
>>
>>> On May 5, 2015, at 3:58 AM, Dmitriy Setrakyan <[hidden email]>
>> wrote:
>>>
>>> I have noticed a Jira filed to make IGFS mountable:
>>> https://issues.apache.org/jira/browse/IGNITE-854
>>>
>>> The integration proposed is via Fuse framework which would allow us mount
>>> IGFS just like a regular Linux file system.
>>>
>>> This seems like a pretty big undertaking. Even though it sounds pretty
>> cool
>>> in theory, it would be interesting to find out form the community if
>>> everyone else finds it useful and maybe suggest some potential use cases
>>> for it.
>>>
>>> Also, seems like Fuse framework is not being actively developed. Last
>>> release came out in July, 2013. How safe is it to use this framework?
>>>
>>> D.
>>
>>

[Konstantin Boudnik-2]

Hadoop is using Fuse to expose HDFS into native Linux environment. It works, I
guess, although I don't know how widely it is used. The whole thing is about
4KLOC + 7KLOC of libhdfs (all in C). While not overly terrible, it is a
significant undertaking indeed.

Having FUSE connector to IGFS is an interesting proposition. I wonder if the
usecases discussed in the thread were covered by simply having a secondary
filesystem configured? Say, being able to sit on top of CEPH would solve all
these needs as CEPH is a kernel-level filesystem. As such, one can work with
it the same way as with any other linux FS. Thoughts?

Cos

On Mon, May 04, 2015 at 06:58PM, Dmitriy Setrakyan wrote:

> I have noticed a Jira filed to make IGFS mountable:
> https://issues.apache.org/jira/browse/IGNITE-854
>
> The integration proposed is via Fuse framework which would allow us mount
> IGFS just like a regular Linux file system.
>
> This seems like a pretty big undertaking. Even though it sounds pretty cool
> in theory, it would be interesting to find out form the community if
> everyone else finds it useful and maybe suggest some potential use cases
> for it.
>
> Also, seems like Fuse framework is not being actively developed. Last
> release came out in July, 2013. How safe is it to use this framework?
>
> D.

[Roman Shaposhnik]

Another alternative for exporting a filesystem is to do it
via java-base NFS server thread. There are a few I
used in the past and there's also the one in HDFS which
may be possible to leverage.

Thanks,
Roman.

On Tue, May 5, 2015 at 12:36 PM, Konstantin Boudnik <[hidden email]> wrote:

> Hadoop is using Fuse to expose HDFS into native Linux environment. It works, I
> guess, although I don't know how widely it is used. The whole thing is about
> 4KLOC + 7KLOC of libhdfs (all in C). While not overly terrible, it is a
> significant undertaking indeed.
>
> Having FUSE connector to IGFS is an interesting proposition. I wonder if the
> usecases discussed in the thread were covered by simply having a secondary
> filesystem configured? Say, being able to sit on top of CEPH would solve all
> these needs as CEPH is a kernel-level filesystem. As such, one can work with
> it the same way as with any other linux FS. Thoughts?
>
> Cos
>
> On Mon, May 04, 2015 at 06:58PM, Dmitriy Setrakyan wrote:
>> I have noticed a Jira filed to make IGFS mountable:
>> https://issues.apache.org/jira/browse/IGNITE-854
>>
>> The integration proposed is via Fuse framework which would allow us mount
>> IGFS just like a regular Linux file system.
>>
>> This seems like a pretty big undertaking. Even though it sounds pretty cool
>> in theory, it would be interesting to find out form the community if
>> everyone else finds it useful and maybe suggest some potential use cases
>> for it.
>>
>> Also, seems like Fuse framework is not being actively developed. Last
>> release came out in July, 2013. How safe is it to use this framework?
>>
>> D.

[dsetrakyan]

On Tue, May 5, 2015 at 12:36 PM, Konstantin Boudnik <[hidden email]> wrote:

> Hadoop is using Fuse to expose HDFS into native Linux environment. It
> works, I
> guess, although I don't know how widely it is used. The whole thing is
> about
> 4KLOC + 7KLOC of libhdfs (all in C). While not overly terrible, it is a
> significant undertaking indeed.
>

Hm... Is Hadoop using Fuse for only HDFS, or for any file system that
implements FileSystem interface? If it is the latter, then IGFS fits right
in into that architecture.

>
> Having FUSE connector to IGFS is an interesting proposition. I wonder if
> the
> usecases discussed in the thread were covered by simply having a secondary
> filesystem configured? Say, being able to sit on top of CEPH would solve
> all
> these needs as CEPH is a kernel-level filesystem. As such, one can work
> with
> it the same way as with any other linux FS. Thoughts?
>
> Cos
>
> On Mon, May 04, 2015 at 06:58PM, Dmitriy Setrakyan wrote:
> > I have noticed a Jira filed to make IGFS mountable:
> > https://issues.apache.org/jira/browse/IGNITE-854
> >
> > The integration proposed is via Fuse framework which would allow us mount
> > IGFS just like a regular Linux file system.
> >
> > This seems like a pretty big undertaking. Even though it sounds pretty
> cool
> > in theory, it would be interesting to find out form the community if
> > everyone else finds it useful and maybe suggest some potential use cases
> > for it.
> >
> > Also, seems like Fuse framework is not being actively developed. Last
> > release came out in July, 2013. How safe is it to use this framework?
> >
> > D.
>

[Konstantin Boudnik-2]

On Tue, May 05, 2015 at 04:49PM, Dmitriy Setrakyan wrote:

> On Tue, May 5, 2015 at 12:36 PM, Konstantin Boudnik <[hidden email]> wrote:
>
> > Hadoop is using Fuse to expose HDFS into native Linux environment. It
> > works, I
> > guess, although I don't know how widely it is used. The whole thing is
> > about
> > 4KLOC + 7KLOC of libhdfs (all in C). While not overly terrible, it is a
> > significant undertaking indeed.
> >
>
> Hm... Is Hadoop using Fuse for only HDFS, or for any file system that
> implements FileSystem interface? If it is the latter, then IGFS fits right
> in into that architecture.

I think it is HDFS specific, but might be wrong on this account.

> > Having FUSE connector to IGFS is an interesting proposition. I wonder if
> > the
> > usecases discussed in the thread were covered by simply having a secondary
> > filesystem configured? Say, being able to sit on top of CEPH would solve
> > all
> > these needs as CEPH is a kernel-level filesystem. As such, one can work
> > with
> > it the same way as with any other linux FS. Thoughts?
> >
> > Cos
> >
> > On Mon, May 04, 2015 at 06:58PM, Dmitriy Setrakyan wrote:
> > > I have noticed a Jira filed to make IGFS mountable:
> > > https://issues.apache.org/jira/browse/IGNITE-854
> > >
> > > The integration proposed is via Fuse framework which would allow us mount
> > > IGFS just like a regular Linux file system.
> > >
> > > This seems like a pretty big undertaking. Even though it sounds pretty
> > cool
> > > in theory, it would be interesting to find out form the community if
> > > everyone else finds it useful and maybe suggest some potential use cases
> > > for it.
> > >
> > > Also, seems like Fuse framework is not being actively developed. Last
> > > release came out in July, 2013. How safe is it to use this framework?
> > >
> > > D.
> >

[Roman Shaposhnik]

On Tue, May 5, 2015 at 5:00 PM, Konstantin Boudnik <[hidden email]> wrote:

> On Tue, May 05, 2015 at 04:49PM, Dmitriy Setrakyan wrote:
>> On Tue, May 5, 2015 at 12:36 PM, Konstantin Boudnik <[hidden email]> wrote:
>>
>> > Hadoop is using Fuse to expose HDFS into native Linux environment. It
>> > works, I
>> > guess, although I don't know how widely it is used. The whole thing is
>> > about
>> > 4KLOC + 7KLOC of libhdfs (all in C). While not overly terrible, it is a
>> > significant undertaking indeed.
>> >
>>
>> Hm... Is Hadoop using Fuse for only HDFS, or for any file system that
>> implements FileSystem interface? If it is the latter, then IGFS fits right
>> in into that architecture.
>
> I think it is HDFS specific, but might be wrong on this account.

Yes it is. But like I said -- NFS implementation could be a way to go.

Thanks,
Roman.

[Konstantin Boudnik-2]

On Tue, May 05, 2015 at 04:59PM, Roman Shaposhnik wrote:

> On Tue, May 5, 2015 at 5:00 PM, Konstantin Boudnik <[hidden email]> wrote:
> > On Tue, May 05, 2015 at 04:49PM, Dmitriy Setrakyan wrote:
> >> On Tue, May 5, 2015 at 12:36 PM, Konstantin Boudnik <[hidden email]> wrote:
> >>
> >> > Hadoop is using Fuse to expose HDFS into native Linux environment. It
> >> > works, I
> >> > guess, although I don't know how widely it is used. The whole thing is
> >> > about
> >> > 4KLOC + 7KLOC of libhdfs (all in C). While not overly terrible, it is a
> >> > significant undertaking indeed.
> >> >
> >>
> >> Hm... Is Hadoop using Fuse for only HDFS, or for any file system that
> >> implements FileSystem interface? If it is the latter, then IGFS fits right
> >> in into that architecture.
> >
> > I think it is HDFS specific, but might be wrong on this account.
>
> Yes it is. But like I said -- NFS implementation could be a way to go.

I believe that Alfresco-like which Marko has referred earlier. I am still not
sure what this feature will solve, so I will seat back and observe this thread
for a while.

    Cos

[Ivan V.]

+1 to Dmitry.
We should implement a feature only if we have clear understanding of its
usecases & benefits it gives to the users.
--ivan

On Wed, May 6, 2015 at 3:04 AM, Konstantin Boudnik <[hidden email]> wrote:

> On Tue, May 05, 2015 at 04:59PM, Roman Shaposhnik wrote:
> > On Tue, May 5, 2015 at 5:00 PM, Konstantin Boudnik <[hidden email]>
> wrote:
> > > On Tue, May 05, 2015 at 04:49PM, Dmitriy Setrakyan wrote:
> > >> On Tue, May 5, 2015 at 12:36 PM, Konstantin Boudnik <[hidden email]>
> wrote:
> > >>
> > >> > Hadoop is using Fuse to expose HDFS into native Linux environment.
> It
> > >> > works, I
> > >> > guess, although I don't know how widely it is used. The whole thing
> is
> > >> > about
> > >> > 4KLOC + 7KLOC of libhdfs (all in C). While not overly terrible, it
> is a
> > >> > significant undertaking indeed.
> > >> >
> > >>
> > >> Hm... Is Hadoop using Fuse for only HDFS, or for any file system that
> > >> implements FileSystem interface? If it is the latter, then IGFS fits
> right
> > >> in into that architecture.
> > >
> > > I think it is HDFS specific, but might be wrong on this account.
> >
> > Yes it is. But like I said -- NFS implementation could be a way to go.
>
> I believe that Alfresco-like which Marko has referred earlier. I am still
> not
> sure what this feature will solve, so I will seat back and observe this
> thread
> for a while.
>
>     Cos
>
>

[Konstantin Boudnik-2]

Perhaps, it might be suffucuent if there's enough interest in the community
and someone is willing to step up and contribute the implementation (and
maintain it, of course)

Cos

On Thu, May 07, 2015 at 04:06PM, Ivan Veselovskiy wrote:

> +1 to Dmitry.
> We should implement a feature only if we have clear understanding of its
> usecases & benefits it gives to the users.
> --ivan
>
> On Wed, May 6, 2015 at 3:04 AM, Konstantin Boudnik <[hidden email]> wrote:
>
> > On Tue, May 05, 2015 at 04:59PM, Roman Shaposhnik wrote:
> > > On Tue, May 5, 2015 at 5:00 PM, Konstantin Boudnik <[hidden email]>
> > wrote:
> > > > On Tue, May 05, 2015 at 04:49PM, Dmitriy Setrakyan wrote:
> > > >> On Tue, May 5, 2015 at 12:36 PM, Konstantin Boudnik <[hidden email]>
> > wrote:
> > > >>
> > > >> > Hadoop is using Fuse to expose HDFS into native Linux environment.
> > It
> > > >> > works, I
> > > >> > guess, although I don't know how widely it is used. The whole thing
> > is
> > > >> > about
> > > >> > 4KLOC + 7KLOC of libhdfs (all in C). While not overly terrible, it
> > is a
> > > >> > significant undertaking indeed.
> > > >> >
> > > >>
> > > >> Hm... Is Hadoop using Fuse for only HDFS, or for any file system that
> > > >> implements FileSystem interface? If it is the latter, then IGFS fits
> > right
> > > >> in into that architecture.
> > > >
> > > > I think it is HDFS specific, but might be wrong on this account.
> > >
> > > Yes it is. But like I said -- NFS implementation could be a way to go.
> >
> > I believe that Alfresco-like which Marko has referred earlier. I am still
> > not
> > sure what this feature will solve, so I will seat back and observe this
> > thread
> > for a while.
> >
> >     Cos
> >
> >

[nivanov]

I think the GlusterFS example proved that distributed file systems generate
significant interest in the user community. GlusterFS used low-level FUSE
APIs to implement its mountability (and gain its well known performance).
Without a native mounting - the file system cannot really be a called a
file system but rather a proprietary API based on file/block access.

I think this feature will only increase Ignite adoption.

--
Nikita Ivanov
​GridGain founder.​



On Thu, May 7, 2015 at 11:44 AM, Konstantin Boudnik <[hidden email]> wrote:

> Perhaps, it might be suffucuent if there's enough interest in the community
> and someone is willing to step up and contribute the implementation (and
> maintain it, of course)
>
> Cos
>
> On Thu, May 07, 2015 at 04:06PM, Ivan Veselovskiy wrote:
> > +1 to Dmitry.
> > We should implement a feature only if we have clear understanding of its
> > usecases & benefits it gives to the users.
> > --ivan
> >
> > On Wed, May 6, 2015 at 3:04 AM, Konstantin Boudnik <[hidden email]>
> wrote:
> >
> > > On Tue, May 05, 2015 at 04:59PM, Roman Shaposhnik wrote:
> > > > On Tue, May 5, 2015 at 5:00 PM, Konstantin Boudnik <[hidden email]>
> > > wrote:
> > > > > On Tue, May 05, 2015 at 04:49PM, Dmitriy Setrakyan wrote:
> > > > >> On Tue, May 5, 2015 at 12:36 PM, Konstantin Boudnik <
> [hidden email]>
> > > wrote:
> > > > >>
> > > > >> > Hadoop is using Fuse to expose HDFS into native Linux
> environment.
> > > It
> > > > >> > works, I
> > > > >> > guess, although I don't know how widely it is used. The whole
> thing
> > > is
> > > > >> > about
> > > > >> > 4KLOC + 7KLOC of libhdfs (all in C). While not overly terrible,
> it
> > > is a
> > > > >> > significant undertaking indeed.
> > > > >> >
> > > > >>
> > > > >> Hm... Is Hadoop using Fuse for only HDFS, or for any file system
> that
> > > > >> implements FileSystem interface? If it is the latter, then IGFS
> fits
> > > right
> > > > >> in into that architecture.
> > > > >
> > > > > I think it is HDFS specific, but might be wrong on this account.
> > > >
> > > > Yes it is. But like I said -- NFS implementation could be a way to
> go.
> > >
> > > I believe that Alfresco-like which Marko has referred earlier. I am
> still
> > > not
> > > sure what this feature will solve, so I will seat back and observe this
> > > thread
> > > for a while.
> > >
> > >     Cos
> > >
> > >
>

[Konstantin Boudnik-2]

On Thu, May 07, 2015 at 01:36PM, Nikita Ivanov wrote:
> I think the GlusterFS example proved that distributed file systems generate
> significant interest in the user community. GlusterFS used low-level FUSE
> APIs to implement its mountability (and gain its well known performance).
> Without a native mounting - the file system cannot really be a called a
> file system but rather a proprietary API based on file/block access.
>
> I think this feature will only increase Ignite adoption.

It is possible! I wonder if something simple and quick can be done to see if
it picks up the interest in further contributions from the community at large?

Cos

> On Thu, May 7, 2015 at 11:44 AM, Konstantin Boudnik <[hidden email]> wrote:
>
> > Perhaps, it might be suffucuent if there's enough interest in the community
> > and someone is willing to step up and contribute the implementation (and
> > maintain it, of course)
> >
> > Cos
> >
> > On Thu, May 07, 2015 at 04:06PM, Ivan Veselovskiy wrote:
> > > +1 to Dmitry.
> > > We should implement a feature only if we have clear understanding of its
> > > usecases & benefits it gives to the users.
> > > --ivan
> > >
> > > On Wed, May 6, 2015 at 3:04 AM, Konstantin Boudnik <[hidden email]>
> > wrote:
> > >
> > > > On Tue, May 05, 2015 at 04:59PM, Roman Shaposhnik wrote:
> > > > > On Tue, May 5, 2015 at 5:00 PM, Konstantin Boudnik <[hidden email]>
> > > > wrote:
> > > > > > On Tue, May 05, 2015 at 04:49PM, Dmitriy Setrakyan wrote:
> > > > > >> On Tue, May 5, 2015 at 12:36 PM, Konstantin Boudnik <
> > [hidden email]>
> > > > wrote:
> > > > > >>
> > > > > >> > Hadoop is using Fuse to expose HDFS into native Linux
> > environment.
> > > > It
> > > > > >> > works, I
> > > > > >> > guess, although I don't know how widely it is used. The whole
> > thing
> > > > is
> > > > > >> > about
> > > > > >> > 4KLOC + 7KLOC of libhdfs (all in C). While not overly terrible,
> > it
> > > > is a
> > > > > >> > significant undertaking indeed.
> > > > > >> >
> > > > > >>
> > > > > >> Hm... Is Hadoop using Fuse for only HDFS, or for any file system
> > that
> > > > > >> implements FileSystem interface? If it is the latter, then IGFS
> > fits
> > > > right
> > > > > >> in into that architecture.
> > > > > >
> > > > > > I think it is HDFS specific, but might be wrong on this account.
> > > > >
> > > > > Yes it is. But like I said -- NFS implementation could be a way to
> > go.
> > > >
> > > > I believe that Alfresco-like which Marko has referred earlier. I am
> > still
> > > > not
> > > > sure what this feature will solve, so I will seat back and observe this
> > > > thread
> > > > for a while.
> > > >
> > > >     Cos
> > > >
> > > >
> >