flink-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Daniel Warneke (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (FLINK-1025) Improve BLOB Service
Date Mon, 14 Jul 2014 20:02:04 GMT

    [ https://issues.apache.org/jira/browse/FLINK-1025?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14061134#comment-14061134
] 

Daniel Warneke commented on FLINK-1025:
---------------------------------------

I’m not emotionally attached to the current design, so just let me know when I shall change
something. However, this is the rationale behind the current version:

1. The hash sum computation is indeed primarily useful for distributing jar files. Since we
cannot expect developers to change the names of their jar files across code changes, the computation
of the hash sum is crucial to guarantee the deployment of the correct user code version. Without
the hash sum computation, we essentially could not locally cache files on the TaskManagers
anymore.

I guess we could debate the overhead of the hash sum computation for hours if we wanted to.
I didn’t know you planned to stress the BLOB service so heavily that it could become a potential
bottleneck. If the service is critical to the performance of the system, it could in fact
be improved quite a bit. If you only plan to push a few MB around, I argue that the overhead
is negligible.

2. The decision not to delete BLOBs after job completion is a consequence of the content-addressable
design. Different jobs might refer to the same file. Therefore it is not safe to delete a
file after a job is finished, unless you start counting references. For now, the BLOB storage
is wiped when the JobManager/TaskManager is shut down. I thought it was good enough to start
with.

3.  I’m perfectly fine with letting the OS chose free ports for the BLOB service.

4. For a general InputStream, you cannot assume to know the length of the incoming content
upfront. Typically, you just read the stream until you get a -1. Unfortunately, for the remote
side to read a -1, the local side has to close the stream. For a socket, however, this operation
closes the entire TCP connection, so it is not possible for the remote side to respond to
the received data anymore. Therefore, the current version transmits the length of the data
that follows per buffer.

5. The singleton implementation is mostly for convenience. There are essentially two parts
in the code where you would have to pass a reference to the BLOB service in a very cumbersome
manner otherwise (the JobGraph and LibraryCacheManager). The current implementation should
work in pseudo clusters as long as the JobManager is instantiated before the TaskManager,
but I’d be open to change the design.


> Improve BLOB Service
> --------------------
>
>                 Key: FLINK-1025
>                 URL: https://issues.apache.org/jira/browse/FLINK-1025
>             Project: Flink
>          Issue Type: Improvement
>          Components: JobManager
>    Affects Versions: 0.6-incubating
>            Reporter: Stephan Ewen
>             Fix For: 0.6-incubating
>
>
> I like the idea of making it transparent where the blob service runs, so the code on
the server/client side is agnostic to that.
> The current merged code is in https://github.com/StephanEwen/incubator-flink/commits/blobservice
> Local tests pass, I am trying distributed tests now.
> There are a few suggestions for improvements:
>  - Since the all the resources are bound to a job or session, it makes sense to make
all puts/gets relative to a jobId (becoming session id) and to have a cleanup hook that delete
all resources associated with that job.
>  - The BLOB service has hardwired to compute a message digest for the contents, and to
use that as the key. While it may make sense for jar files (cached libraries), for many cases
in the future, that will be unnecessary and impose only overhead. I would vote to make this
optional and allow just UUIDs for keys. An example is for the taskmanager to put a part of
an intermediate result into the blob store, for the client to pick it up.
>  - At most points, we have started moving away from configured ports, because of configuration
overhead and collisions in setups, where multiple instances end up on one machine. The latter
happens actually frequently with YARN. I would suggest to have the JM open a port dynamically
for the BlobService (similar as in TaskManager#getAvailablePort() ). RPC calls to figure out
this configuration need to happen only once between client/JM and TM/JM. We can stomach that
overhead ;-)
>  - The write method does not write the length a single time, but "per buffer". Why is
it done that way? The array-based methods know the length up front, and when the contents
comes from an input stream, I think we know the length as well (for files: filesize, for network:
sent up front).
>  - I am personally in favor of moving away from static singleton registries. They tend
to cause trouble during testing, pseudo cluster modes (multiple workers within one JVM). How
hard is it to have a BlobService at the TaskManager / JobManager that we can pass as references
to points where it is needed.



--
This message was sent by Atlassian JIRA
(v6.2#6252)

Mime
View raw message