distributedlog-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Gerrit Sundaram <gerritsunda...@gmail.com>
Subject Re: [DISCUSS] DL Stream Operation Primitives
Date Thu, 17 Nov 2016 09:37:44 GMT
On Tue, Nov 15, 2016 at 2:14 AM, Sijie Guo <sijie@apache.org> wrote:

> On Sat, Nov 12, 2016 at 2:30 AM, Gerrit Sundaram <gerritsundaram@gmail.com
> >
> wrote:
>
> > On Fri, Nov 11, 2016 at 1:09 PM, Sijie Guo <sijie@apache.org> wrote:
> >
> > > I liked this topic. A better name might be 'stream storage primitives',
> > as
> > > we treat DL as a stream storage. Comments inline.
> > >
> > > On Wed, Nov 9, 2016 at 3:09 AM, Gerrit Sundaram <
> > gerritsundaram@gmail.com>
> > > wrote:
> > >
> > > > As what Sijie suggested in the other email thread, I started this
> email
> > > > thread for discussing the stream operation primitives.
> > > >
> > > > The stream operations that I am aware of that DL supports are
> > > >
> > > > * Open a distributedlog stream
> > > > * Delete a distributedlog stream
> > > > * List all the distributedlog streams under a namespace
> > > >
> > >
> > > Are you also looking for listing streams under a 'sub-namespace' - (or
> > > streams have common prefix)? (Based on my understanding on your
> proposal,
> > > you might need this for a filesystem-like API?)
> > >
> >
> > Yes. However it seems like DL is more designed with flat namespace with
> > just streams.
>
>
> Ah, yes. The original thought is to tight a namespace to a user or an
> application. Under a namespace, application can manage the streams by their
> own. So that's why it was designed with a flat namespace.
>
>
> > There is no concept about 'sub-namespace'. Although I
> > probably can hack it by just naming the stream names in a filesystem
> > path-like way.
> >
> > However I am still curious do you guys want to introduce any sort of
> naming
> > hierarchy in the naming within a namespace. For example, can you have a
> > 'StreamSet', which is a set of streams? (like in filesystem, a directory
> > has a list of children). If you have similar hierarchical, it definitely
> > will simply my work.
> >
>
> In the write proxy, we have a similar concept like 'StreamSet' to group
> some physical DL streams into one virtual stream. However that was mostly
> used for exporting metrics for grouped virtual streams. We don't quite
> emphasize the concept of 'virtual stream' in DL. As we tended to let the
> application decide what the virtual stream looks like.
>
> However, for metadata organization and management, it might make sense to
> think of such hierarchy.
>
> What do you have in your mind about 'StreamSet'? Can you explain a little
> more?


I was thinking a group of streams that might be used for same application
but store different parts of data. It is like the 'virtual' stream that you
mentioned.

- Gerrit


>
> >
> >
> > >
> > >
> > > > * Seal a distributedlog stream
> > > > * Truncate a distributedlog stream
> > > >
> > >
> > > Just to clarify this, the 'truncate' in DL is to trim the head of the
> > > stream not the tail.
> > > The 'truncate' in filesystem world is to a size of precisely *length*
> > > bytes, it is truncating the tail.
> > >
> > > Make sure we clarified it and are on same page.
> > >
> >
> > Yes, we are on the same page.
> >
> >
> > >
> > >
> > > >
> > > > I am looking for a more filesystem-like API. for example,
> > > >
> > > > * Get the status/attributes of a stream (like stat in filesystem)
> > > >
> > >
> > > +1 for stream status/attributes. I think we might actually already have
> > > this in DL. since in kestrel, we use that for storing customized
> > metadata.
> > > It might make sense to formalize it into 'stream status'.
> > >
> >
> > Gotcha.
> >
> >
> > >
> > >
> > > > * Rename a stream
> > > >
> > >
> > > we've talked about this for a while. +1.
> > >
> > >
> > > > * Symlink a stream
> > >
> > >
> > > Symlink a stream is probably easy to do. +1 we've thought about that
> for
> > > having the flexibility to move stream between different storage
> backend.
> > > Symlink would help this.
> > >
> > > But a more fundamental thought here is symlinks for log segments. So
> > when a
> > > symlinked stream is deleted, the underneath log segments might not be
> > > deleted until its link count decreased to zero.
> > >
> > >
> > >
> > > >
> > > > Another operations that I can think of might be useful.
> > > >
> > > > * Split/Fork a stream (it can be useful for dynamic data
> partitioning)
> > > >
> > >
> > >
> > >
> > > Split and fork a stream sounds interesting. But it sounds like a more
> > > high-level feature rather than storage primitives. Actually, it might
> be
> > a
> > > good separate discussion feature.
> > >
> > >
> > >
> > >
> > > > * Merge/Concat streams
> > > >
> > >
> > >
> > > I think there is already one outstanding jira for concatenating two DL
> > > streams. Jia and Arvind are working on that.
> > >
> > > https://issues.apache.org/jira/browse/DL-46
> >
> >
> > I will watch that lira.
> >
> >
> > >
> > >
> > >
> > >
> > > >
> > > > The above operations are based on my knowledge about DL. Feel free to
> > add
> > > > more.
> > >
> > >
> > > >
> > > > - Gerrit
> > > >
> > >
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message