On Mon, Aug 26, 2019, at 05:23, Magnus Edenhill wrote:
> Great KIP as always, Colin!
>
> Some comments:
>
> > If the flexible versions are not specified, it is assumed that all
> versions are flexible.
>
> This is ambiguous, if a protocol-generator is pointed to an older Kafka
> protocol specification
> it can't know if the lack of flexibleVersions field means they're all
> flexible, or flexibleVersions
> is not supported at all.
> I recommend requiring the flexibleVersions field if there are flexible
> versions.
Hi Magnus,
Thanks for taking a look.
The issue that I was trying to avoid here is people creating new files and
inadvertently forgetting to set flexibleVersions. Maybe a better way to do it
would be to make flexibleVersions a required field, though. I will change the
KIP to specify that as a requirement.
>
>
> > They are serialized in ascending order of their tag.
>
> I'm not opposed to it, but what's the reasoning for requiring the tag list
> to be ordered?
> And if it is a requirement, make the wording stronger ("required"),
> otherwise remove it.
>
Good question. The reason is that I didn't want to create a situation where
there were multiple different ways to serialize the same data. By having a
single canonical way of serializing it, we can compare two serialized buffers
and know if they represent the same data or different data without
deserializing them. This is very powerful, I think. For example, we can look
at the checksum of two RPCs and know that if they are different, at least
something is different in those RPCs.
>
> > All requests and responses will end with a tagged field buffer. If there
> are no tagged fields, this will only be a single zero byte.
>
> Since the tagged fields can (or most likely will) contain information that
> has bearing on the information in the request or response, it'd be better if
> the tagged fields were inserted before the request/response body.
> A streaming protocol parser would otherwise have to perform two passes of
> the req/resp (worst case).
Hmm. I'm not sure I follow. Each tagged field will be serialized directly
after the struct that contains it. There's no need for two passes when
reading, right? When you read a struct, some fields come first, and some come
later, but that doesn't require two passes.
The rationale for putting the tagged fields after the structure was that we
must do this for the request and response header, and it's simpler to be
consistent than to do something different in the other cases.
>
>
> Addtional comments:
>
> 1)
> With request/response tags we can now introduce request-level errors,
> allowing the broker
> to inform the client why a request was rejected before closing down the
> connection.
> E.g., an empty response body with tag 0 as the error code and tag 1 as the
> error
Yeah... I thought about this a bit as well. We could add a generic response
error as a tagged field in the response header. However, I think the details
of this are complex enough that it's probably worth its own follow-on KIP.
>
> 2)
> The protocol generator should error out if there are duplicate tags for a
> context,
> or if a context has tagged fields but no flexibleVersions, to help catch
> such errors early.
For the first part, the Jackson library takes care of worrying about duplicate
tags. For the second part, I agree, it should be an error if there are tagged
fields but no flexibleVersions specified.
>
> 3)
> A semi-related ask:
> there's currently some non-JSON comments in the protocol spec files,
> typically version history.
> If the information in the comment is important enough to be in the spec
> file, it should be in
> a proper JSON field (e.g., about), which makes sure it ends up in the
> generated protocol docs instead of being filtered out.
>
Yeah, we could consider reorganizing this a bit. This seems like a separate
issue from tagged fields, though.
best,
Colin
>
> /Magnus
>
>
> Den mån 19 aug. 2019 kl 19:00 skrev Colin McCabe <cmcc...@apache.org>:
>
> > Hi Satish,
> >
> > I wasn't originally going to propose supporting the struct type, although
> > perhaps we could consider it.
> >
> > In general, supporting a struct containing an array has the same
> > serialization issues as just supporting the array.
> >
> > Probably, we should just have a two-pass serialization mechanism where we
> > calculate lengths first and then write out bytes. If we do that, we can
> > avoid having any restrictions on tagged fields vs. regular fields. I will
> > take a look at how complex this would be.
> >
> > best,
> > Colin
> >
> >
> > On Sun, Aug 18, 2019, at 22:27, Satish Duggana wrote:
> > > Please read struct type as a complex record type in my earlier mail.
> > > The complex type seems to be defined as Schema[1] in the protocol
> > > types.
> > >
> > > 1.
> > >
> > https://github.com/apache/kafka/blob/trunk/clients/src/main/java/org/apache/kafka/common/protocol/types/Schema.java#L27
> > >
> > >
> > > On Mon, Aug 19, 2019 at 9:46 AM Satish Duggana <satish.dugg...@gmail.com>
> > wrote:
> > > >
> > > > Sorry! Colin, I may not have been clear in my earlier query about
> > > > optional field type restriction. It is mentioned in one of your
> > > > replies "optional fields are serialized starting with their total
> > > > length". This brings the question of whether optional fields support
> > > > struct types (with or without array values). It seems struct types are
> > > > currently not serialized with total length. I may be missing something
> > > > here.
> > > >
> > > > Thanks,
> > > > Satish.
> > > >
> > > >
> > > > On Wed, Aug 14, 2019 at 8:03 AM Satish Duggana <
> > satish.dugg...@gmail.com> wrote:
> > > > >
> > > > > Hi Colin,
> > > > > Thanks for the KIP. Optional fields and var length encoding support
> > is a great
> > > > > improvement for the protocol.
> > > > >
> > > > > >>Optional fields can have any type, except that they cannot be
> > arrays.
> > > > > Note that the restriction against having tagged arrays is just to
> > simplify
> > > > > serialization. We can relax this restriction in the future without
> > changing
> > > > > the protocol on the wire.
> > > > >
> > > > > Can an Optional field have a struct type which internally contains
> > an array
> > > > > field at any level?
> > > > >
> > > > > Thanks,
> > > > > Satish.
> > > > >
> > > > >
> > > > >
> > > > > On Tue, Aug 13, 2019 at 11:49 PM David Jacot <dja...@confluent.io>
> > wrote:
> > > > > >
> > > > > > Hi Colin,
> > > > > >
> > > > > > Thank you for the KIP! Things are well explained!. It is huge
> > improvement
> > > > > > for the Kafka protocol. I have few comments on the proposal:
> > > > > >
> > > > > > 1. The interleaved tag/length header sounds like a great
> > optimisation as it
> > > > > > would be shorter on average. The downside, as
> > > > > > you already pointed out, is that it makes the decoding and the
> > specs more
> > > > > > complex. Personally, I would also favour using two
> > > > > > vaints in this particular case to keep things simple.
> > > > > >
> > > > > > 2. As discussed, I wonder if it would make sense to extend to KIP
> > to also
> > > > > > support optional fields in the Record Header. I think
> > > > > > that it could be interesting to have such capability for common
> > fields
> > > > > > across all the requests or responses (e.g. tracing id).
> > > > > >
> > > > > > Regards,
> > > > > > David
> > > > > >
> > > > > >
> > > > > >
> > > > > > On Tue, Aug 13, 2019 at 10:00 AM Jason Gustafson <
> > ja...@confluent.io> wrote:
> > > > > >
> > > > > > > > Right, I was planning on doing exactly that for all the
> > auto-generated
> > > > > > > RPCs. For the manual RPCs, it would be a lot of work. It’s
> > probably a
> > > > > > > better use of time to convert the manual ones to auto gen first
> > (with the
> > > > > > > possible exception of Fetch/Produce, where the ROI may be higher
> > for the
> > > > > > > manual work)
> > > > > > >
> > > > > > > Yeah, that makes sense. Maybe we can include the version bump
> > for all RPCs
> > > > > > > in this KIP, but we can implement it lazily as the protocols are
> > converted.
> > > > > > >
> > > > > > > -Jason
> > > > > > >
> > > > > > > On Mon, Aug 12, 2019 at 7:16 PM Colin McCabe <cmcc...@apache.org>
> > wrote:
> > > > > > >
> > > > > > > > On Mon, Aug 12, 2019, at 11:22, Jason Gustafson wrote:
> > > > > > > > > Hi Colin,
> > > > > > > > >
> > > > > > > > > Thanks for the KIP! This is a significant improvement. One
> > of my
> > > > > > > personal
> > > > > > > > > interests in this proposal is solving the compatibility
> > problems we
> > > > > > > have
> > > > > > > > > with the internal schemas used to define consumer offsets and
> > > > > > > transaction
> > > > > > > > > metadata. Currently we have to guard schema bumps with the
> > inter-broker
> > > > > > > > > protocol format. Once the format is bumped, there is no way
> > to
> > > > > > > downgrade.
> > > > > > > > > By fixing this, we can potentially begin using the new
> > schemas before
> > > > > > > the
> > > > > > > > > IBP is bumped while still allowing downgrade.
> > > > > > > > >
> > > > > > > > > There are a surprising number of other situations we have
> > encountered
> > > > > > > > this
> > > > > > > > > sort of problem. We have hacked around it in special cases
> > by allowing
> > > > > > > > > nullable fields to the end of the schema, but this is not
> > really an
> > > > > > > > > extensible approach. I'm looking forward to having a better
> > option.
> > > > > > > >
> > > > > > > > Yeah, this problem keeps coming up.
> > > > > > > >
> > > > > > > > >
> > > > > > > > > With that said, I have a couple questions on the proposal:
> > > > > > > > >
> > > > > > > > > 1. For each request API, we need one version bump to begin
> > support for
> > > > > > > > > "flexible versions." Until then, we won't have the option of
> > using
> > > > > > > tagged
> > > > > > > > > fields even if the broker knows how to handle them. Does it
> > make sense
> > > > > > > to
> > > > > > > > > go ahead and do a universal bump of each request API now so
> > that we'll
> > > > > > > > have
> > > > > > > > > this option going forward?
> > > > > > > >
> > > > > > > > Right, I was planning on doing exactly that for all the
> > auto-generated
> > > > > > > > RPCs. For the manual RPCs, it would be a lot of work. It’s
> > probably a
> > > > > > > > better use of time to convert the manual ones to auto gen
> > first (with the
> > > > > > > > possible exception of Fetch/Produce, where the ROI may be
> > higher for the
> > > > > > > > manual work)
> > > > > > > >
> > > > > > > > > 2. The alternating length/tag header encoding lets us save a
> > byte in
> > > > > > > the
> > > > > > > > > common case. The downside is that it's a bit more complex to
> > specify.
> > > > > > > It
> > > > > > > > > also has some extra cost if the length exceeds the tag
> > substantially.
> > > > > > > > > Basically we'd have to pad the tag, right? I think I'm
> > wondering if we
> > > > > > > > > should just bite the bullet and use two varints instead.
> > > > > > > >
> > > > > > > > That’s a fair point. It would be shorter on average, but worse
> > for some
> > > > > > > > exceptional cases. Also, the decoding would be more complex,
> > which might
> > > > > > > be
> > > > > > > > a good reason to go for just having two varints. Yeah, let’s
> > simplify.
> > > > > > > >
> > > > > > > > Regards,
> > > > > > > > Colin
> > > > > > > >
> > > > > > > > >
> > > > > > > > > Thanks,
> > > > > > > > > Jason
> > > > > > > > >
> > > > > > > > >
> > > > > > > > > On Fri, Aug 9, 2019 at 4:31 PM Colin McCabe <
> > cmcc...@apache.org>
> > > > > > > wrote:
> > > > > > > > >
> > > > > > > > > > Hi all,
> > > > > > > > > >
> > > > > > > > > > I've made some updates to this KIP. Specifically, I wanted
> > to avoid
> > > > > > > > > > including escape bytes in the serialization format, since
> > it was too
> > > > > > > > > > complex. Also, I think this is a good opportunity to slim
> > down our
> > > > > > > > > > variable length fields.
> > > > > > > > > >
> > > > > > > > > > best,
> > > > > > > > > > Colin
> > > > > > > > > >
> > > > > > > > > >
> > > > > > > > > > On Thu, Jul 11, 2019, at 20:52, Colin McCabe wrote:
> > > > > > > > > > > On Tue, Jul 9, 2019, at 15:29, Jose Armando Garcia
> > Sancio wrote:
> > > > > > > > > > > > Thanks Colin for the KIP. For my own edification why
> > are we doing
> > > > > > > > this
> > > > > > > > > > > > "Optional fields can have any type, except for an
> > array of
> > > > > > > > > > structures."?
> > > > > > > > > > > > Why can't we have an array of structures?
> > > > > > > > > > >
> > > > > > > > > > > Optional fields are serialized starting with their total
> > length.
> > > > > > > This
> > > > > > > > > > > is straightforward to calculate for primitive fields
> > like INT32,
> > > > > > > (or
> > > > > > > > > > > even an array of INT32), but more difficult to calculate
> > for an
> > > > > > > array
> > > > > > > > > > > of structures. Basically, we'd have to do a two-pass
> > serialization
> > > > > > > > > > > where we first calculate the lengths of everything, and
> > then write
> > > > > > > it
> > > > > > > > > > > out.
> > > > > > > > > > >
> > > > > > > > > > > The nice thing about this KIP is that there's nothing in
> > the
> > > > > > > protocol
> > > > > > > > > > > stopping us from adding support for this feature in the
> > future. We
> > > > > > > > > > > wouldn't have to really change the protocol at all to
> > add support.
> > > > > > > > But
> > > > > > > > > > > we'd have to change a lot of serialization code. Given
> > almost all
> > > > > > > of
> > > > > > > > > > > our use-cases for optional fields are adding an extra
> > field here or
> > > > > > > > > > > there, it seems reasonable not to support it for right
> > now.
> > > > > > > > > > >
> > > > > > > > > > > best,
> > > > > > > > > > > Colin
> > > > > > > > > > >
> > > > > > > > > > > >
> > > > > > > > > > > > --
> > > > > > > > > > > > -Jose
> > > > > > > > > > > >
> > > > > > > > > > >
> > > > > > > > > >
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > >
> >
>
