kafka-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From radai <radai.rosenbl...@gmail.com>
Subject Re: [DISCUSS] KIP-82 - Add Record Headers
Date Wed, 09 Nov 2016 00:23:33 GMT
+1 for sean's document. it covers pretty much all the trade-offs and
provides concrete figures to argue about :-)
(nit-picking - used the same xkcd twice, also trove has been superceded for
purposes of high performance collections: look at
https://github.com/leventov/Koloboke)

so to sum up the string vs int debate:

performance - you can do 140k ops/sec _per thread_ with string headers. you
could do x2-3 better with ints. there's no arguing the relative diff
between the two, there's only the question of whether or not _the rest of
kafka_ operates fast enough to care. if we want to make choices solely
based on performance we need ints. if we are willing to settle/compromise
for a nicer (to some) API than strings are good enough for the current
state of affairs.

message size - with batching and compression it comes down to a ~5%
difference (internal testing, not in the doc. maybe would help adding if
this becomes a point of contention?). this means it wont really affect
kafka in "throughput mode" (large, compressed batches). in "low latency"
mode (meaning less/no batching and compression) the difference can be
extreme (it'll easily be an order of magnitude with small payloads like
stock ticks and header keys of the form
"com.acme.infraTeam.kafka.hiMom.auditPlugin"). we have a few such topics at
linkedin where actual payloads are ~2 ints and are eclipsed by our in-house
audit "header" which is why we liked ints to begin with.

"ease of use" - strings would probably still require _some_ degree of
partitioning by convention (imagine if everyone used the key "infra"...)
but its very intuitive for java devs to do anyway (reverse-domain is
ingrained into java developers at a young age :-) ). also most java devs
find Map<String, whatever> more intuitive than Map<Integer, whatever> -
probably because of other text-based protocols like http. ints would
require a number registry. if you think number registries are hard just
look at the wiki page for KIPs (specifically the number for next available
KIP) and think again - we are probably talking about the same volume of
requests. also this would only be "required" (good citizenship, more like)
if you want to publish your plugin for others to use. within your org do
whatever you want - just know that if you use [some "reserved" range] and a
future kafka update breaks it its your problem. RTFM.

personally im in favor of ints.

having said that (and like nacho) I will settle if int vs string remains
the only obstacle to this.

On Tue, Nov 8, 2016 at 3:53 PM, Nacho Solis <nsolis@linkedin.com.invalid>
wrote:

> I think it's well known I've been pushing for ints (and I could switch to
> 16 bit shorts if pressed).
>
> - efficient (space)
> - efficient (processing)
> - easily partitionable
>
>
> However, if the only thing that is keeping us from adopting headers is the
> use of strings vs ints as keys, then I would cave in and accept strings. If
> we do so, I would like to limit string keys to 128 bytes in length.  This
> way 1) I could use a 3 letter string if I wanted (effectively using 4 total
> bytes), 2) limit overall impact of possible keys (don't really want people
> to send a 16K header string key).
>
> Nacho
>
>
> On Tue, Nov 8, 2016 at 3:35 PM, Gwen Shapira <gwen@confluent.io> wrote:
>
> > Forgot to mention: Thank you for quantifying the trade-off - it is
> > helpful and important regardless of what we end up deciding.
> >
> > On Tue, Nov 8, 2016 at 3:12 PM, Sean McCauliff
> > <smccauliff@linkedin.com.invalid> wrote:
> > > On Tue, Nov 8, 2016 at 2:15 PM, Gwen Shapira <gwen@confluent.io>
> wrote:
> > >
> > >> Since Kafka specifically targets high-throughput, low-latency
> > >> use-cases, I don't think we should trade them off that easily.
> > >>
> > >
> > > I find these kind of design goals not to be really helpful unless it's
> > > quantified in someway.  Because it's always possible to argue against
> > > something as either being not performant or just an implementation
> > detail.
> > >
> > > This is a single threaded benchmarks so all the measurements are per
> > > thread.
> > >
> > > For 1M messages/s/thread  if header keys are int and you had even a
> > single
> > > header key, value pair then it's still about 2^-2 microseconds which
> > means
> > > you only have another 0.75 microseconds to do everything else you want
> to
> > > do with a message (1M messages/s means 1 micro second per message).
> With
> > > string header keys there is still 0.5 micro seconds to process a
> message.
> > >
> > >
> > >
> > > I love strings as much as the next guy (we had them in Flume), but I
> > >> was convinced by Magnus/Michael/Radai that strings don't actually have
> > >> strong benefits as opposed to ints (you'll need a string registry
> > >> anyway - otherwise, how will you know what does the "profile_id"
> > >> header refers to?) and I want to keep closer to our original design
> > >> goals for Kafka.
> > >>
> > >
> > > "confluent.profile_id"
> > >
> > >
> > >>
> > >> If someone likes strings in the headers and doesn't do millions of
> > >> messages a sec, they probably have lots of other systems they can use
> > >> instead.
> > >>
> > >
> > > None of them will scale like Kafka.  Horizontal scaling is still good.
> > >
> > >
> > >>
> > >>
> > >> On Tue, Nov 8, 2016 at 1:22 PM, Sean McCauliff
> > >> <smccauliff@linkedin.com.invalid> wrote:
> > >> > +1 for String keys.
> > >> >
> > >> > I've been doing some bechmarking and it seems like the speedup for
> > using
> > >> > integer keys is about 2-5 depending on the length of the strings and
> > what
> > >> > collections are being used.  The overall amount of time spent
> parsing
> > a
> > >> set
> > >> > of header key, value pairs probably does not matter unless you are
> > >> getting
> > >> > close to 1M messages per consumer.  In which case probably don't use
> > >> > headers.  There is also the option to use very short strings; some
> > that
> > >> are
> > >> > even shorter than integers.
> > >> >
> > >> > Partitioning the string key space will be easier than partitioning
> an
> > >> > integer key space. We won't need a global registry.  Kafka
> internally
> > can
> > >> > reserve some prefix like "_" as its namespace.  Everyone else can
> use
> > >> their
> > >> > company or project name as namespace prefix and life should be good.
> > >> >
> > >> > Here's the link to some of the benchmarking info:
> > >> > https://docs.google.com/document/d/1tfT-
> > 6SZdnKOLyWGDH82kS30PnUkmgb7nPL
> > >> dw6p65pAI/edit?usp=sharing
> > >> >
> > >> >
> > >> >
> > >> > --
> > >> > Sean McCauliff
> > >> > Staff Software Engineer
> > >> > Kafka
> > >> >
> > >> > smccauliff@linkedin.com
> > >> > linkedin.com/in/sean-mccauliff-b563192
> > >> >
> > >> > On Mon, Nov 7, 2016 at 11:51 PM, Michael Pearce <
> > Michael.Pearce@ig.com>
> > >> > wrote:
> > >> >
> > >> >> +1 on this slimmer version of our proposal
> > >> >>
> > >> >> I def think the Id space we can reduce from the proposed
> > int32(4bytes)
> > >> >> down to int16(2bytes) it saves on space and as headers we wouldn't
> > >> expect
> > >> >> the number of headers being used concurrently being that high.
> > >> >>
> > >> >> I would wonder if we should make the value byte array length still
> > int32
> > >> >> though as This is the standard Max array length in Java saying
that
> > it
> > >> is a
> > >> >> header and I guess limiting the size is sensible and would work
for
> > all
> > >> the
> > >> >> use cases we have in mind so happy with limiting this.
> > >> >>
> > >> >> Do people generally concur on Magnus's slimmer version? Anyone
see
> > any
> > >> >> issues if we moved from int32 to int16?
> > >> >>
> > >> >> Re configurable ids per plugin over a global registry also would
> work
> > >> for
> > >> >> us.  As such if this has better concensus over the proposed global
> > >> registry
> > >> >> I'd be happy to change that.
> > >> >>
> > >> >> I was already sold on ints over strings for keys ;)
> > >> >>
> > >> >> Cheers
> > >> >> Mike
> > >> >>
> > >> >> ________________________________________
> > >> >> From: Magnus Edenhill <magnus@edenhill.se>
> > >> >> Sent: Monday, November 7, 2016 10:10:21 PM
> > >> >> To: dev@kafka.apache.org
> > >> >> Subject: Re: [DISCUSS] KIP-82 - Add Record Headers
> > >> >>
> > >> >> Hi,
> > >> >>
> > >> >> I'm +1 for adding generic message headers, but I do share the
> > concerns
> > >> >> previously aired on this thread and during the KIP meeting.
> > >> >>
> > >> >> So let me propose a slimmer alternative that does not require
any
> > sort
> > >> of
> > >> >> global header registry, does not affect broker performance or
> > >> operations,
> > >> >> and adds as little overhead as possible.
> > >> >>
> > >> >>
> > >> >> Message
> > >> >> ------------
> > >> >> The protocol Message type is extended with a Headers array consting
> > of
> > >> >> Tags, where a Tag is defined as:
> > >> >>    int16 Id
> > >> >>    int16 Len              // binary_data length
> > >> >>    binary_data[Len]  // opaque binary data
> > >> >>
> > >> >>
> > >> >> Ids
> > >> >> ---
> > >> >> The Id space is not centrally managed, so whenever an application
> > needs
> > >> to
> > >> >> add headers, or use an eco-system plugin that does, its Id
> allocation
> > >> will
> > >> >> need to be manually configured.
> > >> >> This moves the allocation concern from the global space down to
> > >> >> organization level and avoids the risk for id conflicts.
> > >> >> Example pseudo-config for some app:
> > >> >>     sometrackerplugin.tag.sourcev3.id=1000
> > >> >>     dbthing.tag.tablename.id=1001
> > >> >>     myschemareg.tag.schemaname.id=1002
> > >> >>     myschemareg.tag.schemaversion.id=1003
> > >> >>
> > >> >>
> > >> >> Each header-writing or header-reading plugin must provide means
> > >> (typically
> > >> >> through configuration) to specify the tag for each header it uses.
> > >> Defaults
> > >> >> should be avoided.
> > >> >> A consumer silently ignores tags it does not have a mapping for
> > (since
> > >> the
> > >> >> binary_data can't be parsed without knowing what it is).
> > >> >>
> > >> >> Id range 0..999 is reserved for future use by the broker and must
> > not be
> > >> >> used by plugins.
> > >> >>
> > >> >>
> > >> >>
> > >> >> Broker
> > >> >> ---------
> > >> >> The broker does not process the tags (other than the standard
> > protocol
> > >> >> syntax verification), it simply stores and forwards them as opaque
> > data.
> > >> >>
> > >> >> Standard message translation (removal of Headers) kicks in for
> older
> > >> >> clients.
> > >> >>
> > >> >>
> > >> >> Why not string ids?
> > >> >> -------------------------
> > >> >> String ids might seem like a good idea, but:
> > >> >>  * does not really solve uniqueness
> > >> >>  * consumes a lot of space (2 byte string length + string, per
> > header)
> > >> to
> > >> >> be meaningful
> > >> >>  * doesn't really say anything how to parse the tag's data, so
it
> is
> > in
> > >> >> effect useless on its own.
> > >> >>
> > >> >>
> > >> >> Regards,
> > >> >> Magnus
> > >> >>
> > >> >>
> > >> >>
> > >> >>
> > >> >> 2016-11-07 18:32 GMT+01:00 Michael Pearce <Michael.Pearce@ig.com>:
> > >> >>
> > >> >> > Hi Roger,
> > >> >> >
> > >> >> > Thanks for the support.
> > >> >> >
> > >> >> > I think the key thing is to have a common key space to make
an
> > >> ecosystem,
> > >> >> > there does have to be some level of contract for people to
play
> > >> nicely.
> > >> >> >
> > >> >> > Having map<String, byte[]> or as per current proposed
in kip of
> > >> having a
> > >> >> > numerical key space of  map<int, byte[]> is a level
of the
> contract
> > >> that
> > >> >> > most people would expect.
> > >> >> >
> > >> >> > I think the example in a previous comment someone else made
> > linking to
> > >> >> AWS
> > >> >> > blog and also implemented api where originally they didn’t
have a
> > >> header
> > >> >> > space but not they do, where keys are uniform but the value
can
> be
> > >> >> string,
> > >> >> > int, anything is a good example.
> > >> >> >
> > >> >> > Having a custom MetadataSerializer is something we had played
> with,
> > >> but
> > >> >> > discounted the idea, as if you wanted everyone to work the
same
> > way in
> > >> >> the
> > >> >> > ecosystem, having to have this also customizable makes it
a bit
> > >> harder.
> > >> >> > Think about making the whole message record custom serializable,
> > this
> > >> >> would
> > >> >> > make it fairly tricky (though it would not be impossible)
to have
> > made
> > >> >> work
> > >> >> > nicely. Having the value customizable we thought is a reasonable
> > >> tradeoff
> > >> >> > here of flexibility over contract of interaction between
> different
> > >> >> parties.
> > >> >> >
> > >> >> > Is there a particular case or benefit of having serialization
> > >> >> customizable
> > >> >> > that you have in mind?
> > >> >> >
> > >> >> > Saying this it is obviously something that could be implemented,
> if
> > >> there
> > >> >> > is a need. If we did go this avenue I think a defaulted
> serializer
> > >> >> > implementation should exist so for the 80:20 rule, people
can
> just
> > >> have
> > >> >> the
> > >> >> > broker and clients get default behavior.
> > >> >> >
> > >> >> > Cheers
> > >> >> > Mike
> > >> >> >
> > >> >> > On 11/6/16, 5:25 PM, "radai" <radai.rosenblatt@gmail.com>
wrote:
> > >> >> >
> > >> >> >     making header _key_ serialization configurable potentially
> > >> undermines
> > >> >> > the
> > >> >> >     board usefulness of the feature (any point along the
path
> must
> > be
> > >> >> able
> > >> >> > to
> > >> >> >     read the header keys. the values may be whatever and
require
> > more
> > >> >> > intimate
> > >> >> >     knowledge of the code that produced specific headers,
but
> keys
> > >> should
> > >> >> > be
> > >> >> >     universally readable).
> > >> >> >
> > >> >> >     it would also make it hard to write really portable plugins
-
> > say
> > >> i
> > >> >> > wrote a
> > >> >> >     large message splitter/combiner - if i rely on key
> > "largeMessage"
> > >> and
> > >> >> >     values of the form "1/20" someone who uses (contrived
> example)
> > >> >> > Map<Byte[],
> > >> >> >     Double> wouldnt be able to re-use my code.
> > >> >> >
> > >> >> >     not the end of a the world within an organization, but
> > >> problematic if
> > >> >> > you
> > >> >> >     want to enable an ecosystem
> > >> >> >
> > >> >> >     On Thu, Nov 3, 2016 at 2:04 PM, Roger Hoover <
> > >> roger.hoover@gmail.com
> > >> >> >
> > >> >> > wrote:
> > >> >> >
> > >> >> >     >  As others have laid out, I see strong reasons for
a common
> > >> message
> > >> >> >     > metadata structure for the Kafka ecosystem.  In
particular,
> > I've
> > >> >> > seen that
> > >> >> >     > even within a single organization, infrastructure
teams
> often
> > >> own
> > >> >> the
> > >> >> >     > message metadata while application teams own the
> > >> application-level
> > >> >> > data
> > >> >> >     > format.  Allowing metadata and content to have different
> > >> structure
> > >> >> > and
> > >> >> >     > evolve separately is very helpful for this.  Also,
I think
> > >> there's
> > >> >> a
> > >> >> > lot of
> > >> >> >     > value to having a common metadata structure shared
across
> the
> > >> Kafka
> > >> >> >     > ecosystem so that tools which leverage metadata
can more
> > easily
> > >> be
> > >> >> > shared
> > >> >> >     > across organizations and integrated together.
> > >> >> >     >
> > >> >> >     > The question is, where does the metadata structure
belong?
> > >> Here's
> > >> >> > my take:
> > >> >> >     >
> > >> >> >     > We change the Kafka wire and on-disk format to from
a (key,
> > >> value)
> > >> >> > model to
> > >> >> >     > a (key, metadata, value) model where all three are
byte
> > arrays
> > >> from
> > >> >> > the
> > >> >> >     > brokers point of view.  The primary reason for this
is that
> > it
> > >> >> > provides a
> > >> >> >     > backward compatible migration path forward.  Producers
can
> > start
> > >> >> > populating
> > >> >> >     > metadata fields before all consumers understand
the
> metadata
> > >> >> > structure.
> > >> >> >     > For people who already have custom envelope structures,
> they
> > can
> > >> >> > populate
> > >> >> >     > their existing structure and the new structure for
a while
> as
> > >> they
> > >> >> > make the
> > >> >> >     > transition.
> > >> >> >     >
> > >> >> >     > We could stop there and let the clients plug in
a
> > KeySerializer,
> > >> >> >     > MetadataSerializer, and ValueSerializer but I think
it is
> > also
> > >> be
> > >> >> > useful to
> > >> >> >     > have a default MetadataSerializer that implements
a
> key-value
> > >> model
> > >> >> > similar
> > >> >> >     > to AMQP or HTTP headers.  Or we could go even further
and
> > >> >> prescribe a
> > >> >> >     > Map<String, byte[]> or Map<String, String>
data model for
> > >> headers
> > >> >> in
> > >> >> > the
> > >> >> >     > clients (while still allowing custom serialization
of the
> > header
> > >> >> data
> > >> >> >     > model).
> > >> >> >     >
> > >> >> >     > I think this would address Radai's concerns:
> > >> >> >     > 1. All client code would not need to be updated
to know
> about
> > >> the
> > >> >> >     > container.
> > >> >> >     > 2. Middleware friendly clients would have a standard
header
> > data
> > >> >> > model to
> > >> >> >     > work with.
> > >> >> >     > 3. KIP is required both b/c of broker changes and
because
> of
> > >> client
> > >> >> > API
> > >> >> >     > changes.
> > >> >> >     >
> > >> >> >     > Cheers,
> > >> >> >     >
> > >> >> >     > Roger
> > >> >> >     >
> > >> >> >     >
> > >> >> >     > On Wed, Nov 2, 2016 at 4:38 PM, radai <
> > >> radai.rosenblatt@gmail.com>
> > >> >> > wrote:
> > >> >> >     >
> > >> >> >     > > my biggest issues with a "standard" wrapper
format:
> > >> >> >     > >
> > >> >> >     > > 1. _ALL_ client _CODE_ (as opposed to kafka
lib version)
> > must
> > >> be
> > >> >> > updated
> > >> >> >     > to
> > >> >> >     > > know about the container, because any old naive
code
> > trying to
> > >> >> > directly
> > >> >> >     > > deserialize its own payload would keel over
and die (it
> > needs
> > >> to
> > >> >> > know to
> > >> >> >     > > deserialize a container, and then dig in there
for its
> > >> payload).
> > >> >> >     > > 2. in order to write middleware-friendly clients
that
> > utilize
> > >> >> such
> > >> >> > a
> > >> >> >     > > container one would basically have to write
their own
> > >> >> > producer/consumer
> > >> >> >     > API
> > >> >> >     > > on top of the open source kafka one.
> > >> >> >     > > 3. if you were going to go with a wrapper format
you
> really
> > >> dont
> > >> >> > need to
> > >> >> >     > > bother with a kip (just open source your own
client stack
> > >> from #2
> > >> >> > above
> > >> >> >     > so
> > >> >> >     > > others could stop re-inventing it)
> > >> >> >     > >
> > >> >> >     > > On Wed, Nov 2, 2016 at 4:25 PM, James Cheng
<
> > >> >> wushujames@gmail.com>
> > >> >> >     > wrote:
> > >> >> >     > >
> > >> >> >     > > > How exactly would this work? Or maybe
that's out of
> scope
> > >> for
> > >> >> > this
> > >> >> >     > email.
> > >> >> >     > >
> > >> >> >     >
> > >> >> >
> > >> >> >
> > >> >> > The information contained in this email is strictly confidential
> > and
> > >> for
> > >> >> > the use of the addressee only, unless otherwise indicated.
If you
> > are
> > >> not
> > >> >> > the intended recipient, please do not read, copy, use or
disclose
> > to
> > >> >> others
> > >> >> > this message or any attachment. Please also notify the sender
by
> > >> replying
> > >> >> > to this email or by telephone (+44(020 7896 0011) and then
delete
> > the
> > >> >> email
> > >> >> > and any copies of it. Opinions, conclusion (etc) that do
not
> > relate to
> > >> >> the
> > >> >> > official business of this company shall be understood as
neither
> > given
> > >> >> nor
> > >> >> > endorsed by it. IG is a trading name of IG Markets Limited
(a
> > company
> > >> >> > registered in England and Wales, company number 04008957)
and IG
> > Index
> > >> >> > Limited (a company registered in England and Wales, company
> number
> > >> >> > 01190902). Registered address at Cannon Bridge House, 25
Dowgate
> > Hill,
> > >> >> > London EC4R 2YA. Both IG Markets Limited (register number
195355)
> > and
> > >> IG
> > >> >> > Index Limited (register number 114059) are authorised and
> > regulated by
> > >> >> the
> > >> >> > Financial Conduct Authority.
> > >> >> >
> > >> >> The information contained in this email is strictly confidential
> and
> > for
> > >> >> the use of the addressee only, unless otherwise indicated. If
you
> are
> > >> not
> > >> >> the intended recipient, please do not read, copy, use or disclose
> to
> > >> others
> > >> >> this message or any attachment. Please also notify the sender
by
> > >> replying
> > >> >> to this email or by telephone (+44(020 7896 0011) and then delete
> the
> > >> email
> > >> >> and any copies of it. Opinions, conclusion (etc) that do not relate
> > to
> > >> the
> > >> >> official business of this company shall be understood as neither
> > given
> > >> nor
> > >> >> endorsed by it. IG is a trading name of IG Markets Limited (a
> company
> > >> >> registered in England and Wales, company number 04008957) and
IG
> > Index
> > >> >> Limited (a company registered in England and Wales, company number
> > >> >> 01190902). Registered address at Cannon Bridge House, 25 Dowgate
> > Hill,
> > >> >> London EC4R 2YA. Both IG Markets Limited (register number 195355)
> > and IG
> > >> >> Index Limited (register number 114059) are authorised and regulated
> > by
> > >> the
> > >> >> Financial Conduct Authority.
> > >> >>
> > >>
> > >>
> > >>
> > >> --
> > >> Gwen Shapira
> > >> Product Manager | Confluent
> > >> 650.450.2760 | @gwenshap
> > >> Follow us: Twitter | blog
> > >>
> >
> >
> >
> > --
> > Gwen Shapira
> > Product Manager | Confluent
> > 650.450.2760 | @gwenshap
> > Follow us: Twitter | blog
> >
>
>
>
> --
> Nacho (Ignacio) Solis
> Kafka
> nsolis@linkedin.com
>

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