kafka-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Guozhang Wang <wangg...@gmail.com>
Subject Re: [DISCUSS] KIP-280: Enhanced log compaction
Date Fri, 01 Jun 2018 18:51:36 GMT
Hello Luis,

Please feel free to continue on the voting process as there seems be no
further comments on this thread (I have synced with Jun and Ismael
separately offline and they are in consent with the approach to add the
fields in offset map for all cases).

We can still continue on reviewing the PR while voting on the thread so
that it can get in earlier into trunk for the next release.



Guozhang


On Mon, May 28, 2018 at 11:04 AM, Matthias J. Sax <matthias@confluent.io>
wrote:

> Luis,
>
> this week is feature freeze for the upcoming 2.0 release and most people
> focus on getting their PR merged. Thus, this and the next week (until
> code freeze) KIPs for 2.1 are not a high priority for most people.
>
> Please bear with us. Thanks for your understanding.
>
>
> -Matthias
>
> On 5/28/18 5:21 AM, Luís Cabral wrote:
> >  Hi Guozhang,
> >
> > It doesn't look like there will be much feedback here.
> > Is it alright if I just update the spec back to a standardized behaviour
> and move this along?
> >
> > Cheers,Luis
> >     On Thursday, May 24, 2018, 11:20:01 AM GMT+2, Luis Cabral <
> luis_cabral@yahoo.com> wrote:
> >
> >  Hi Jun / Ismael,
> >
> > Any chance to get your opinion on this?
> > Thanks in advance!
> >
> > Regards,
> > Luís
> >
> >> On 22 May 2018, at 17:30, Guozhang Wang <wangguoz@gmail.com> wrote:
> >>
> >> Hello Luís,
> >>
> >> While reviewing your PR I realized my previous calculation on the memory
> >> usage was incorrect: in fact, in the current implementation, each entry
> in
> >> the memory-bounded cache is 16 (default MD5 hash digest length) + 8
> (long
> >> type) = 24 bytes, and if we add the long-typed version value it is 32
> >> bytes. I.e. each entry will be increased by 33%, not doubling.
> >>
> >> After redoing the math I'm bit leaning towards just adding this entry
> for
> >> all cases rather than treating timestamp differently with others (sorry
> for
> >> being back and forth, but I just want to make sure we've got a good
> balance
> >> between efficiency and semantics consistency). I've also chatted with
> Jun
> >> and Ismael about this (cc'ed), and maybe you guys can chime in here as
> well.
> >>
> >>
> >> Guozhang
> >>
> >>
> >> On Tue, May 22, 2018 at 6:45 AM, Luís Cabral
> <Luis_Cabral@yahoo.com.invalid>
> >> wrote:
> >>
> >>> Hi Matthias / Guozhang,
> >>>
> >>> Were the questions clarified?
> >>> Please feel free to add more feedback, otherwise it would be nice to
> move
> >>> this topic onwards 😊
> >>>
> >>> Kind Regards,
> >>> Luís Cabral
> >>>
> >>> From: Guozhang Wang
> >>> Sent: 09 May 2018 20:00
> >>> To: dev@kafka.apache.org
> >>> Subject: Re: [DISCUSS] KIP-280: Enhanced log compaction
> >>>
> >>> I have thought about being consistency in strategy v.s. practical
> concerns
> >>> about storage convenience to its impact on compaction effectiveness.
> >>>
> >>> The different between timestamp and the header key-value pairs is that
> for
> >>> the latter, as I mentioned before, "it is arguably out of Kafka's
> control,
> >>> and indeed users may (mistakenly) generate many records with the same
> key
> >>> and the same header value." So giving up tie breakers may result in
> very
> >>> poor compaction effectiveness when it happens, while for timestamps the
> >>> likelihood of this is considered very small.
> >>>
> >>>
> >>> Guozhang
> >>>
> >>>
> >>> On Sun, May 6, 2018 at 8:55 PM, Matthias J. Sax <matthias@confluent.io
> >
> >>> wrote:
> >>>
> >>>> Thanks.
> >>>>
> >>>> To reverse the question: if this argument holds, why does it not apply
> >>>> to the case when the header key is used as compaction attribute?
> >>>>
> >>>> I am not against keeping both records in case timestamps are equal,
> but
> >>>> shouldn't we apply the same strategy for all cases and don't use
> offset
> >>>> as tie-breaker at all?
> >>>>
> >>>>
> >>>> -Matthias
> >>>>
> >>>>> On 5/6/18 8:47 PM, Guozhang Wang wrote:
> >>>>> Hello Matthias,
> >>>>>
> >>>>> The related discussion was in the PR:
> >>>>> https://github.com/apache/kafka/pull/4822#discussion_r184588037
> >>>>>
> >>>>> The concern is that, to use offset as tie breaker we need to double
> the
> >>>>> entry size of the entry in bounded compaction cache, and hence
> largely
> >>>>> reduce the effectiveness of the compaction itself. Since with
> >>>> milliseconds
> >>>>> timestamp the scenario of ties with the same key is expected to
be
> >>>> small, I
> >>>>> think it would be a reasonable tradeoff to make.
> >>>>>
> >>>>>
> >>>>> Guozhang
> >>>>>
> >>>>> On Sun, May 6, 2018 at 9:37 AM, Matthias J. Sax <
> matthias@confluent.io
> >>>>
> >>>>> wrote:
> >>>>>
> >>>>>> Hi,
> >>>>>>
> >>>>>> I just updated myself on this KIP. One question (maybe it was
> >>> discussed
> >>>>>> and I missed it). What is the motivation to not use the offset
as
> tie
> >>>>>> breaker for the "timestamp" case? Isn't this inconsistent behavior?
> >>>>>>
> >>>>>>
> >>>>>> -Matthias
> >>>>>>
> >>>>>>
> >>>>>>> On 5/2/18 2:07 PM, Guozhang Wang wrote:
> >>>>>>> Hello Luís,
> >>>>>>>
> >>>>>>> Sorry for the late reply.
> >>>>>>>
> >>>>>>> My understanding is that such duplicates will only happen
if the
> >>>>>> non-offset
> >>>>>>> version value, either the timestamp or some long-typed header
key,
> >>> are
> >>>>>> the
> >>>>>>> same (i.e. we cannot break ties).
> >>>>>>>
> >>>>>>> 1. For timestamp, which is in milli-seconds, I think in
practice
> the
> >>>>>>> likelihood of records with the same key and the same milli-sec
> >>>> timestamp
> >>>>>>> are very small. And hence the duplicate amount should be
very
> small.
> >>>>>>>
> >>>>>>> 2. For long-typed header key, it is arguably out of Kafka's
> control,
> >>>> and
> >>>>>>> indeed users may (mistakenly) generate many records with
the same
> key
> >>>> and
> >>>>>>> the same header value.
> >>>>>>>
> >>>>>>>
> >>>>>>> So I'd like to propose a counter-offer: for 1), we still
use only 8
> >>>> bytes
> >>>>>>> and allows for potential duplicates due to ties; for 2)
we use 16
> >>> bytes
> >>>>>> to
> >>>>>>> always break ties. The motivation for distinguishing 1)
and 2), is
> >>> that
> >>>>>> my
> >>>>>>> expectation for 1) would be much common, and hence worth
special
> >>>> handling
> >>>>>>> it to be more effective in cleaning. WDYT?
> >>>>>>>
> >>>>>>>
> >>>>>>> Guozhang
> >>>>>>>
> >>>>>>>
> >>>>>>>
> >>>>>>> On Wed, May 2, 2018 at 2:36 AM, Luís Cabral
> >>>>>> <luis_cabral@yahoo.com.invalid>
> >>>>>>> wrote:
> >>>>>>>
> >>>>>>>> Hi Guozhang,
> >>>>>>>>
> >>>>>>>> Have you managed to have a look at my reply?
> >>>>>>>> How do you feel about this?
> >>>>>>>>
> >>>>>>>> Kind Regards,
> >>>>>>>> Luís Cabral
> >>>>>>>>     On Monday, April 30, 2018, 9:27:15 AM GMT+2, Luís
Cabral <
> >>>>>>>> luis_cabral@yahoo.com> wrote:
> >>>>>>>>
> >>>>>>>>   Hi Guozhang,
> >>>>>>>>
> >>>>>>>> I understand the argument, but this is a hazardous compromise
for
> >>>> using
> >>>>>>>> Kafka as an event store (as is my original intention).
> >>>>>>>>
> >>>>>>>> I expect to have many duplicated messages in Kafka as
the overall
> >>>>>>>> architecture being used allows for the producer to re-send
a fresh
> >>>>>> state of
> >>>>>>>> the backed data into Kafka.Though this scenario is not
common, as
> >>> the
> >>>>>>>> intention is for Kafka to bear the weight of replaying
all the
> >>> records
> >>>>>> for
> >>>>>>>> new consumers, but it will occasionally happen.
> >>>>>>>>
> >>>>>>>> As there are plenty of records which are not updated
frequently,
> >>> this
> >>>>>>>> would leave the topic with a surplus of quite a few
million
> >>> duplicate
> >>>>>>>> records (and increasing every time the above mentioned
function is
> >>>>>> applied).
> >>>>>>>>
> >>>>>>>> I would prefer to have the temporary memory footprint
of 8 bytes
> per
> >>>>>>>> record whenever the compaction is run (only when not
in 'offset'
> >>>> mode),
> >>>>>>>> than allowing for the topic to run into this state.
> >>>>>>>>
> >>>>>>>> What do you think? Is this scenario too specific for
me, or do you
> >>>>>> believe
> >>>>>>>> that it could happen to other clients as well?
> >>>>>>>>
> >>>>>>>> Thanks again for the continued discussion!
> >>>>>>>> Cheers,
> >>>>>>>> Luis    On Friday, April 27, 2018, 8:21:13 PM GMT+2,
Guozhang
> Wang <
> >>>>>>>> wangguoz@gmail.com> wrote:
> >>>>>>>>
> >>>>>>>> Hello Luis,
> >>>>>>>>
> >>>>>>>> When the comparing the version returns `equal`, the
original
> >>> proposal
> >>>>>> is to
> >>>>>>>> use the offset as the tie breaker. My previous comment
is that
> >>>>>>>>
> >>>>>>>> 1) when we build the map calling `put`, if there is
already an
> entry
> >>>> for
> >>>>>>>> the key, compare its stored version, and replace if
the put
> record's
> >>>>>>>> version is "no smaller than" the stored record: this
is because
> when
> >>>>>>>> building the map we are always going from smaller offsets
to
> larger
> >>>>>> ones.
> >>>>>>>>
> >>>>>>>> 2) when making a second pass to determine if each record
should be
> >>>>>> retained
> >>>>>>>> based on the map, we do not try to break the tie if
the map's
> >>> returned
> >>>>>>>> version is the same but always treat it as "keep". In
this case
> when
> >>>> we
> >>>>>> are
> >>>>>>>> comparing a record with itself stored in the offset
map, version
> >>>>>> comparison
> >>>>>>>> would return `equals`. As I mentioned in the PR, one
caveat is
> that
> >>> we
> >>>>>> may
> >>>>>>>> indeed have multiple records with the same key and the
same
> version,
> >>>> but
> >>>>>>>> once a new versioned record is appended it will be deleted.
> >>>>>>>>
> >>>>>>>>
> >>>>>>>> Does that make sense?
> >>>>>>>>
> >>>>>>>> Guozhang
> >>>>>>>>
> >>>>>>>>
> >>>>>>>> On Fri, Apr 27, 2018 at 4:20 AM, Luís Cabral
> >>>>>> <luis_cabral@yahoo.com.invalid
> >>>>>>>>>
> >>>>>>>> wrote:
> >>>>>>>>
> >>>>>>>>> Hi,
> >>>>>>>>>
> >>>>>>>>> I was updating the PR to match the latest decisions
and noticed
> (or
> >>>>>>>>> rather, the integration tests noticed) that without
storing the
> >>>> offset,
> >>>>>>>>> then the cache doesn't know when to keep the record
itself.
> >>>>>>>>>
> >>>>>>>>> This is because, after the cache is populated, all
the records
> are
> >>>>>>>>> compared against the stored ones, so "Record{key:A,offset:1,
> >>>>>> version:1}"
> >>>>>>>>> will compare against itself and be flagged as "don't
keep", since
> >>> we
> >>>>>> only
> >>>>>>>>> compared based on the version and didn't check to
see if the
> offset
> >>>> was
> >>>>>>>> the
> >>>>>>>>> same or not.
> >>>>>>>>>
> >>>>>>>>> This sort of invalidates not storing the offset
in the cache,
> >>>>>>>>> unfortunately, and the binary footprint increases
two-fold when
> >>>>>> "offset"
> >>>>>>>> is
> >>>>>>>>> not used as a compaction strategy.
> >>>>>>>>>
> >>>>>>>>> Guozhang: Is it ok with you if we go back on this
decision and
> >>> leave
> >>>>>> the
> >>>>>>>>> offset as a tie-breaker?
> >>>>>>>>>
> >>>>>>>>>
> >>>>>>>>> Kind Regards,Luis
> >>>>>>>>>
> >>>>>>>>>   On Friday, April 27, 2018, 11:11:55 AM GMT+2,
Luís Cabral
> >>>>>>>>> <luis_cabral@yahoo.com.INVALID> wrote:
> >>>>>>>>>
> >>>>>>>>> Hi,
> >>>>>>>>>
> >>>>>>>>> The KIP is now updated with the results of the byte
array
> >>> discussion.
> >>>>>>>>>
> >>>>>>>>> This is my first contribution to Kafka, so I'm not
sure on what
> the
> >>>>>>>>> processes are. Is it now acceptable to take this
into a vote, or
> >>>>>> should I
> >>>>>>>>> ask for more contributors to join the discussion
first?
> >>>>>>>>>
> >>>>>>>>> Kind Regards,Luis    On Friday, April 27, 2018,
6:12:03 AM GMT+2,
> >>>>>>>> Guozhang
> >>>>>>>>> Wang <wangguoz@gmail.com> wrote:
> >>>>>>>>>
> >>>>>>>>> Hello Luís,
> >>>>>>>>>
> >>>>>>>>>> Offset is an integer? I've only noticed it being
resolved as a
> >>> long
> >>>> so
> >>>>>>>>> far.
> >>>>>>>>>
> >>>>>>>>> You are right, offset is a long.
> >>>>>>>>>
> >>>>>>>>> As for timestamp / other types, I left a comment
in your PR about
> >>>>>>>> handling
> >>>>>>>>> tie breakers.
> >>>>>>>>>
> >>>>>>>>>> Given these arguments, is this point something
that you
> absolutely
> >>>>>> must
> >>>>>>>>> have?
> >>>>>>>>>
> >>>>>>>>> No I do not have a strong use case in mind to go
with arbitrary
> >>> byte
> >>>>>>>>> arrays, was just thinking that if we are going to
enhance log
> >>>>>> compaction
> >>>>>>>>> why not generalize it more :)
> >>>>>>>>>
> >>>>>>>>> Your concern about the memory usage makes sense.
I'm happy to
> take
> >>> my
> >>>>>>>>> suggestion back and enforce only long typed fields.
> >>>>>>>>>
> >>>>>>>>>
> >>>>>>>>> Guozhang
> >>>>>>>>>
> >>>>>>>>>
> >>>>>>>>>
> >>>>>>>>>
> >>>>>>>>>
> >>>>>>>>> On Thu, Apr 26, 2018 at 1:44 AM, Luís Cabral
> >>>>>>>> <luis_cabral@yahoo.com.invalid
> >>>>>>>>>>
> >>>>>>>>> wrote:
> >>>>>>>>>
> >>>>>>>>>> Hi,
> >>>>>>>>>>
> >>>>>>>>>> bq. have a integer typed OffsetMap (for offset)
> >>>>>>>>>>
> >>>>>>>>>> Offset is an integer? I've only noticed it being
resolved as a
> >>> long
> >>>> so
> >>>>>>>>> far.
> >>>>>>>>>>
> >>>>>>>>>>
> >>>>>>>>>> bq. long typed OffsetMap (for timestamp)
> >>>>>>>>>>
> >>>>>>>>>> We would still need to store the offset, as
it is functioning
> as a
> >>>>>>>>>> tie-breaker. Not that this is a big deal, we
can be easily have
> >>> both
> >>>>>>>> (as
> >>>>>>>>>> currently done on the PR).
> >>>>>>>>>>
> >>>>>>>>>>
> >>>>>>>>>> bq. For the byte array typed offset map, we
can use a general
> >>>> hashmap,
> >>>>>>>>>> where the hashmap's CAPACITY will be reasoned
from the given
> "val
> >>>>>>>> memory:
> >>>>>>>>>> Int" parameter
> >>>>>>>>>>
> >>>>>>>>>> If you have a map with 128 byte capacity, then
store a value
> with
> >>> 16
> >>>>>>>>> bytes
> >>>>>>>>>> and another with 32 bytes, how many free slots
do you have left
> in
> >>>>>> this
> >>>>>>>>> map?
> >>>>>>>>>>
> >>>>>>>>>> You can make this work, but I think you would
need to re-design
> >>> the
> >>>>>>>> whole
> >>>>>>>>>> log cleaner approach, which implies changing
some of the already
> >>>>>>>> existing
> >>>>>>>>>> configurations (like "log.cleaner.io.buffer.load.factor").
I
> >>> would
> >>>>>>>>> rather
> >>>>>>>>>> maintain backwards compatibility as much as
possible in this
> KIP,
> >>>> and
> >>>>>>>> if
> >>>>>>>>>> this means that using "foo" / "bar" or "2.1-a"
/ "3.20-b" as
> >>> record
> >>>>>>>>>> versions is not viable, then so be it.
> >>>>>>>>>>
> >>>>>>>>>> Given these arguments, is this point something
that you
> absolutely
> >>>>>> must
> >>>>>>>>>> have? I'm still sort of hoping that you are
just entertaining
> the
> >>>> idea
> >>>>>>>>> and
> >>>>>>>>>> are ok with having a long (now conceded to be
unsigned, so the
> >>> byte
> >>>>>>>>> arrays
> >>>>>>>>>> can be compared directly).
> >>>>>>>>>>
> >>>>>>>>>>
> >>>>>>>>>> Kind Regards,Luis
> >>>>>>>>>>
> >>>>>>>>>
> >>>>>>>>>
> >>>>>>>>>
> >>>>>>>>> --
> >>>>>>>>> -- Guozhang
> >>>>>>>>>
> >>>>>>>>
> >>>>>>>>
> >>>>>>>>
> >>>>>>>> --
> >>>>>>>> -- Guozhang
> >>>>>>>>
> >>>>>>>
> >>>>>>>
> >>>>>>>
> >>>>>>
> >>>>>>
> >>>>>
> >>>>>
> >>>>
> >>>>
> >>>
> >>>
> >>> --
> >>> -- Guozhang
> >>>
> >>>
> >>
> >>
> >> --
> >> -- Guozhang
>
>


-- 
-- Guozhang

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