May 4, 2026 • 35 mins
Tim Berglund talks to Caleb Grillo (Confluent / WarpStream) about his career in data streaming product management. Caleb’s first job: washing windows. Their challenge: reshaping Confluent Cloud’s billing and pioneering diskless Kafka to trade latency for huge cost savings.
SEASON 2
Hosted by Tim Berglund, Adi Polak and Viktor Gamov
Produced and Edited by Noelle Gallagher, Peter Furia and Nurie Mohamed
Music by Coastal Kites
Artwork by Phil Vo
- 🎧 Subscribe to Confluent Developer wherever you listen to podcasts.
- ▶️ Subscribe on YouTube, and hit the 🔔 to catch new episodes.
- 👍 If you enjoyed this, please leave us a rating.
- 🎧 Confluent also has a podcast for tech leaders: "Life Is But A Stream" hosted by our friend, Joseph Morais.
Listen
Watch
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
SPEAKER_00 (00:00):
But we pioneered this idea that you can sacrifice
latency in order to gain hugeamounts of cost savings,
specifically infrastructure costsavings.
I mean, the thing that peopleneed to understand about iceberg
is that it is a it's a tableformat, it's a building block.
Everyone's like, oh, good PMsare right a lot.
You're you're wrong, like almostall the time.
You just have to be wrong in theright direction.
SPEAKER_01 (00:21):
Hey there, everybody, I'm Tim Berglund, and
welcome to Confluent Developer,the podcast where we explore the
fascinating journeys of softwaredevelopers tackling complex
problems.
In this episode, I talked toCaleb Grillo, the staff product
manager for Confluent WarpStream.
Caleb recounts his storystarting as a PM in the early
days of Confluent Cloud, talkingabout billing and a maturing
(00:42):
platform, and then later joiningan exciting young startup,
WarpStream, in late 2023.
WarpStream at the time was busyturning the world of Kafka
upside down, and Caleb tells thewhole story from the perspective
of a data-driven product managerwho is always wrong in
somebody's eyes.
Caleb tells a great story, andnow we get to listen.
(01:17):
Yeah.
Hey, tell us a little bit aboutyourself.
What uh what are you what areyou doing these days?
SPEAKER_00 (01:24):
Yeah, uh so I'm I'm Caleb.
Um I'm our uh product lead forthe Warp Stream product line at
Confluent, uh, which is our uhBYOC product line.
So bring your own cloud.
Um we uh we'll we'll get intoall the details of that, but
yeah, I I run our our productmanagement function um here.
(01:45):
And uh yeah, so that day to day,it's everything's different uh
every day.
Um so no no two days are reallythe same uh for me, but I work I
work across uh our engineeringteam to figure out you know what
we're what we're building andwhether it's the right thing for
our customers and make sure thatwe're prioritizing all the right
(02:06):
things.
Uh but I also do a fair amountof sort of analytics work, um
you know, digging into sort ofyou know how people are using
our product.
Um and I also work on sort ofthe business side, the GTM, um,
making sure that our sort ofcommercials make sense and and
sort of integration with thelarger confluent business and
(02:29):
all that.
So it's a pretty uh that's a lotknee knee deep, mile-wide type
role.
SPEAKER_01 (02:35):
What do you what do you mean analytics?
I thought being a PM was justvibes.
You're saying you use data?
SPEAKER_00 (02:39):
No, no, no, no.
Yeah, we uh so you know, all ofour consumption data and and all
of our usage data goes into umconfluence data warehouse, and
so we're able to sort of um dolike trend analysis on things
like revenue and you know costand make sure that our cost and
margin makes sense and all thatstuff.
SPEAKER_01 (02:58):
Yeah, so it's it's good it's good to have that
because there's I I've saidthere's probably a lot of
reasons why I'm not a productmanager, like you know, maybe
qualification or something, butI think you'd be a great one,
too.
SPEAKER_00 (03:09):
Well, here's the thing though.
SPEAKER_01 (03:10):
Here's the thing when you're a PM, like every
decision you make is to asignificant fraction of your
stakeholders, not just wrong,but like how could you be that
stupid wrong?
Yes.
So you're you're it's this thisprocess of being idiotically
wrong all the time.
In somebody's eye, like 40, 50percent of the people think that
(03:31):
all the time.
I got into a line of work whereI I get like the constant
affirmation that I need, likebeing on stage and stuff like
that.
So I think that's the it'sreally just an inner thing.
You know, I I need people totell me how awesome I am.
You need the strength ofcharacter to do the right thing,
even when they tell you you'rewrong.
SPEAKER_00 (03:47):
Oh you have that.
I think to be to be a reallygreat product person, you have
to be a glutton for punishment.
Yep.
Uh and you have to everyone'slike, oh, good PMs are right a
lot.
You're you're wrong, like almostall the time.
You just have to be wrong in theright direction.
SPEAKER_01 (04:02):
Yeah, and whether in terms of like objective results,
you're you're right or wrong,you you could come up with some
rubric for scoring that.
Somebody thinks you're wrong allthe time.
SPEAKER_00 (04:11):
Yeah, oh yeah, yeah, yeah.
SPEAKER_01 (04:12):
Yeah, you're just a guy, you're a moron all the
time.
And that I would I wouldstruggle.
Uh, what was what was your firstjob?
SPEAKER_00 (04:20):
Yeah, my first job out of out of college was not
um, well, my first job ever wasuh washing was washing windows
um in in high school.
Uh I had uh my one of mybrother's friends, I have an
older brother, one of mybrother's friends had a window
washing business and we'd goaround people's houses and wash
(04:40):
their windows.
That was my first job ever.
It was like a summer job.
Um my first serious job out ofcollege, uh, I worked for an
international development NGO.
Um and so yeah, it was reallyinteresting.
Um, we did a lot of work forlike USAID, the Agency for
International Development, uh,but also for sort of UN agencies
(05:03):
and other things like that.
Um and so, you know, and it wasreally global, like there was
there was there were projects inall kinds of different places
and different differentcountries and continents.
I didn't get to travel verymuch, uh, just you know, because
I I moved on to tech uh before Isort of got into travel-related
work.
But um that role was kind ofinteresting.
(05:24):
I was I was uh working onproposal budgeting.
So, you know, there's like atechnical portion of these, you
know, sort of project proposals,uh, and there's a cost
component.
And so I was working on the costcomponent, which is figuring
out, you know, how much does aToyota land cruiser cost in
Uganda, uh, you know, andputting that into budget
proposals and things like that.
(05:45):
Um so yeah, I mean it it was itwas sort of interesting.
I guess you could I I didn't atthe time think of it this way,
but you could draw someparallels between that and
product management, you know,making sure that all the pieces
fit together and make sense.
Um but uh yeah, I mean that thatwas my first job out of college.
Um yeah.
SPEAKER_01 (06:06):
There you go.
Window washing prepares you forproduct management only if your
customers berate you about whata terrible job you did washing
windows.
Yeah, that's right.
Um too slow.
Too slow.
You washed the wrong window panefirst, and uh the pattern you
used was wrong.
SPEAKER_00 (06:24):
I read in a book somewhere that you were supposed
to do it some other way.
SPEAKER_01 (06:27):
There you go.
This wiki this Wikipedia page.
Um tell us about I think youkind of hinted this, but what's
uh most interesting problem youever saw to date?
SPEAKER_00 (06:38):
Um yeah, so uh well I guess I can start with my my
uh my character arc um becausethat'll that'll sort of explain
my answer.
Uh love that, yeah.
I yeah, so so I I uh I got intoum tech after a couple years of
working for the InternationalDevelopment MBO NGO.
(06:59):
Um I I got into tech uh in sortof a business operations role
and then moved into productmanagement a couple years later.
Um and you know, one of my firstproduct jobs uh was um dealing
with a lot of the streaming dataat uh at an e-commerce site
(07:21):
that's not Amazon.
Uh and so um we were basicallybuilding a data platform for uh
for competitive intelligencedata.
So it was lots of high volumedata um that was being like
scraped from various places onthe internet, uh, basically
building a huge data lake of allthat information.
(07:41):
So teams like you know dataanalytics and data science could
figure out which prod productswere set were gonna sell well
against others and you know,things like that, uh, who which
sites were listing all thesedifferent products, how the
review counts were correlatingwith uh you know, and review
ratings were correlating withlike listing those those uh
(08:03):
products and surfacing them oncompetitor sites, basically
trying to learn what was goingon in the market from external
data sources instead of just ourown internal analytics.
And that uh that pipeline, likethe ingest pipeline, was Kafka.
Um so this was 2017, 2018.
SPEAKER_01 (08:20):
Um that was around the start of my my own Kafka
journey.
SPEAKER_00 (08:25):
Yeah, yeah, yeah, yeah.
Similar uh similar tracks, Iguess.
Um and so you know, I I came tolearn about Confluent when I was
in that role and we solved a lotof interesting problems.
Um, but I I kind of realizedlike, okay, there's like a more
scalable way to do this, like togo to Confluent and work on the
technology under underpinningthis this whole thing.
(08:46):
Um I just got super interestedin that.
And so I I ended up uh movingover to Confluent in May 2019.
Um and when I started, uh we hada sort of a you you'll remember
this, um, we had a sort ofdisjoint uh like product lineup.
There was in the there was acloud product that had gained a
lot of traction, but that cloudproduct was not really what you
(09:08):
would expect from likecommercially from a cloud
product.
Um, meaning like it wasn't likeAWS where you spin up an EC2
instance and you start beingbilled hourly for that instance
and you spin it down and it youstop being billed.
So there was there was like nopossibility of self-serve
anything uh in the in thatversion of our cloud product.
SPEAKER_01 (09:29):
It's it's it's in our defense back then, it's very
hard to make a cloud service.
And so, yeah, that was superprimitive.
Yeah.
SPEAKER_00 (09:37):
And it was it was like it was a proof of concept.
Like it was like uh, you know,will people use a streaming
platform hosted by somebodyelse?
Yes, that was proven uh thatthey would, and it kind of it
went too far.
SPEAKER_01 (09:51):
Uh we got carried away, went crazy.
SPEAKER_00 (09:55):
Yeah, yeah, it got carried away with the with the
proof of concept, and and thenit was time to make it a real
product.
Uh and so um when I came in atComplet, that was kind of the
state.
There was like, you know, thebig cloud business was this um
sort of interesting hybrid oflike a hosted service, um, but
it wasn't really what you'dexpect from like a fully managed
(10:17):
cloud product.
It was sort of just we'll runyour Kafka brokers for you.
SPEAKER_01 (10:21):
Um Kafka on some servers is yeah, yeah, yeah,
yeah, yeah.
SPEAKER_00 (10:26):
And we wanted like the vision for Comple and Cloud
was always to be way more thanthat.
Um servers.
And so yeah, and so so like thethe biggest the the the sort of
biggest blocker that we had,other than sort of the hard
technical things of you knowbuilding a a real cloud service,
um, we had the underpinnings ofthat, but like we were limited
(10:47):
by the commercial model.
Um and so that was sort of thethat was the big problem that
needed to be solved um back inyou know 2019 was like how do we
make the commercial model workum to in order to unlock sort of
the self-serve you know cloudvision of a cloud product that
(11:08):
everybody had.
Um and sort of central to thatwas building out the
consumption-based billingsystem, and so Confluent had a
consumption-based billing systemalready, but it was for their
self-serve cloud product.
So there were two differentproducts.
There was the non-self-serve,serious usage, dedicated
(11:31):
infrastructure hosted brokersproduct.
Yeah, you like it.
And there was another product,yeah, and there was another
product that was kind of justlike, you know, go sign up for
this thing, put in your creditcard, it's self-serve.
That was all multi-tenantclusters.
Um, so like there were all theunderpinnings that we needed to
sort of put everything together,but it was impossible to marry
those two worlds togetherwithout the commercial model
(11:54):
making sense.
SPEAKER_01 (11:54):
Okay.
SPEAKER_00 (11:55):
And so, in order to have a consumption-based
commercial model, you had tohave a consumption-based billing
system that could handle youknow all the different sort of
cluster types you'd need and allthe different infrastructure,
like there's you know, KafkaConnect, there's KSQL.
At the time, Flink wasn't there,but you need it to be sort of
able to support all thesedifferent um products and
(12:17):
different like cluster types,and you need to be able to meter
throughput and storage and likeall the different usage metrics.
Um, and so building out thefirst version of that was kind
of uh welcome to Confluent.
Here's here's this huge project,you need to go figure out how to
do it.
SPEAKER_01 (12:34):
Now, a quick word from our sponsor.
Confluent developer the podcastis brought to you by Confluent
Developer the website, which haseverything you need as a
developer of data streamingsystems.
And it's completely free.
We've got curriculum, hands-onexercises, tutorials, the online
data streaming engineercertification are also free.
(12:54):
A way to find a meetup near you,those are free, everything is
there.
I really want you to besuccessful in your journey as a
data streaming engineer, andthis is the site that has what
you need.
Check it out atdeveloper.confluent.io.
That's developer.confluent.io.
Now back to the show.
Um that was you and I werecoworkers then, but remind me,
(13:17):
were you a PM or an engineer?
What where were you then?
SPEAKER_00 (13:20):
I was a PM.
Yeah, yeah, yeah.
So I I I started I started aconfluent as a PM.
Uh and and so I yeah, that thatwas the first sort of thing to
figure out.
And that involved like it washugely cross-functional, uh,
which is a you know, hard thingin itself.
SPEAKER_01 (13:39):
Always.
SPEAKER_00 (13:39):
Meaning I had to work like with engineering, uh,
but also all the other PMs withrunning all their other
services, and also the salesteam and like to and the
marketing team, like the thewhole way things happened at
Confluent had to shift.
Uh, it wasn't just, oh, let'sjust build a billing system and
(13:59):
build it and they will come.
Like we had to make everythingwork.
The finance team, like we hadweekly meetings with like you
know, 10 different heads of X inthe room trying to figure out
sort of what we were doing.
Uh and so that was difficult.
It was also technicallydifficult because you know, like
(14:21):
any metering, like at the time,there weren't these nice SaaS
tools that you could buy to sortof solve the technical
underpinning of your billingsystem.
Like every company that wantedto do this had to build their
own system.
SPEAKER_01 (14:37):
Yeah.
SPEAKER_00 (14:37):
Right.
Yeah.
Uh and so we had already built,you know, before I got there,
there was there was like thissort of pre-existing metering
system.
Um, but it didn't support, forexample, having multiple cluster
tiers.
We couldn't have like adedicated cluster and a
multi-tenant cluster sittingnext to each other being built.
Uh, so we had to like build thatin.
(14:58):
Um there's also, you know, thisis getting into sort of boring
business stuff, but like there'suh there's a concept of like a
commit or like a minimumcommitment that you know a
customer would make in order tohave sort of a long-term deal
with Confluent.
Um and building that conceptinto the billing system.
(15:18):
Like we had to do that becauseyour usage gets metered, but you
can't just meter it at listprice if somebody's getting a
discount.
So we needed a concept of adiscount.
Like there were like all thesethings we had to model in and
sort of we were changing thewheels on the 747 as it was
coming in for a landing, youknow.
That's the metaphor.
SPEAKER_01 (15:37):
There's customers, customers using the platform at
this point.
SPEAKER_00 (15:40):
Yes, yeah, yeah, yeah, yeah.
That's true.
And then we had to come up witha strategy for migrating people
over and making sure that thepricing made sense, and like,
you know, we're just we'retrying to merge all these worlds
together.
Uh, so it's very, veryinteresting and very, very high
impact, uh, but also very hardum to get it right.
And you know, we didn't, I don'tthink we got it 100% right the
(16:01):
first time, uh, but we we builtit in a way that was like that
it could in theory support allthe future iterations that you
currently see in ConfluentCloud.
Yeah, um, you know, and it'stotally different now.
It's like a completely differentsystem.
Um, the technical side of it waswas also super interesting
because you know the theinfrastructure under the hood is
(16:24):
you know throwing off all sortsof metrics.
There, it's you know, you youcan measure all kinds of stuff.
Um, and figuring out like whatwhat pipelines, like what parts
of the observability pipelinewould be suitable for billing
was kind of a um, you know, thatwas a challenge in itself.
Uh figuring out which metricsmeant the things that they
(16:46):
needed to mean when it comes tobilling, because that's a
different problem thanobservability.
Yes.
Um, yes, I mean it it it was itwas uh it was a big undertaking,
but I I think that it, you know,I I learned a ton.
Uh it's a great, it's a greatintroduction to a company to be
like, hey, can we just likechange everything about it?
SPEAKER_01 (17:02):
Uh yeah.
Totally cross-functionally.
Different agendas, differentmotivation, different incentive
structures, different uhpersonalities, and and and you
make it all work.
SPEAKER_00 (17:14):
Yeah, yeah, yeah.
The technical side of it alsowas was super cool.
Uh, I don't want to minimizethat.
Like that that was um for mevery interesting to figure out
like how to translate all ofthese different requirements,
very specific, like you know uhlike gigabytes of cons of of
(17:36):
rights, like rights to thecluster like means something
specific.
There's like six differentmetrics that you could
potentially think would be youknow used for measuring rights
and what's the right one.
Uh and so you know, there'sthere's stuff like that.
And then you know, um different,like you know, I guess if uh
(17:57):
somebody wants to be able to doper second billing, it's like,
well, actually we want toaggregate things hourly and send
present it hourly to cut likethere there were there were lots
of arguments about like whatspecific things we did and and
we had to figure out like whatwas supportable with the current
platform and what when you havethings like you know bandwidth
limitations and things likethat, that implies uh a
(18:20):
measurement regime and a timegranular.
SPEAKER_01 (18:23):
There's all kinds of things that that have to be true
that are their own engineeringproblems to solve.
And dirty secret in any givensystem that has a thing like
that, you know, that might beover the whole day or something,
you know, that it isn'tnecessarily like an
instantaneous uh thing, it justdepends on what you've built.
(18:45):
And it's it's all its own way.
SPEAKER_00 (18:48):
Yeah, this was something that was interesting
that we always had to explain topeople.
It was like there's like aperiod of time where when you
like log into the UI and youlook at the billing screen,
there's a period of time wherethe amounts can change.
That like broke people's brainswhen you're trying to explain
this to customers.
They were like, and this was youknow, 2019, 2020, you know,
(19:09):
yeah, six years ago.
Uh people would would look atthat statement and be like,
wait, but I can I trust yourbilling system and be like,
yeah, yeah, we have all thesecorrectness checks, and
everything's like everything'sgood.
It's just that there can be somelate-arriving metrics.
Like this is a massivedistributed system of
distributed systems.
Like you can have stuff happen,but don't worry, like your
(19:30):
monthly bill is fine.
SPEAKER_01 (19:31):
Yeah, yeah.
It'll converge, it'll converge.
SPEAKER_00 (19:34):
Yeah, yeah, it'll converge, right?
SPEAKER_01 (19:37):
So you um you ended up fast-forwarding a little bit
at at Warp Stream.
Uh how'd you get there?
unknown (19:46):
Yeah.
SPEAKER_00 (19:46):
Yeah, so uh I had been at Confluent for about five
years.
Um and actually it it came aboutbasically because Rishi, uh the
one of the co-founders ofWarpstream, posted this blog
post, which is now very famous,uh, called Kafka is dead, long
live Kafka.
(20:07):
And all the ideas that they wereputting out there were just very
interesting.
They were basically saying thatlike there's this category of
workload where you know Kafka isa super low latency real-time
system.
There's this category ofworkload that doesn't need to be
super low latency, like it's notextremely latency sensitive.
(20:30):
And I remember this, you know,for years and years.
It was just like it wasn't evena thought that could cross your
mind that you coulddifferentiate a a product on
latency.
It was like, no, it's just fixedconstraint.
It's a real-time system, latencymust be small.
SPEAKER_01 (20:46):
It's um you know, is it is it zero yet?
Well, keep working, you know.
SPEAKER_00 (20:50):
Yeah, exactly.
And it's like it there, therewas never the insight of like
you can differentiate uh on thisdimension that like is a little
bit counterintuitive if you'rejust glancing at it.
But if you think about it for asecond, it's like if you're
running an observabilityplatform for customers, you
know, that's your product.
You your your customer doesn'tnotice if it takes a few hundred
(21:14):
milliseconds for your ingestpipeline to receive data.
Like it does not matter in thegrand scheme of things because
there's this whole pipeline thatneeds to happen that needs to
run to process that data beforeit can ever be displayed on a
graph in your product.
SPEAKER_01 (21:31):
And I I remember, I I think it was the summer of
2020.
Like you could check me on that,but reading that blog post and
just kind of warp stream comingon my radar and realizing
there's 23.
23.
Okay.
Okay.
Um realizing I don't know whatthe the like the demand
(21:51):
elasticity of uh of latencyreally is.
Like how many how many people ifif you could give them half the
price of Or a tenth of theprice, or you know, do that, but
it takes a second.
How many people care?
Like, I didn't know.
And it was just it was that samerevelation.
Like, wow, that could be 90% ofthe market that just doesn't
(22:13):
care.
I and and I still don't think wereally know that, but um it was
very interesting.
unknown (22:19):
Yeah.
SPEAKER_00 (22:20):
Yeah, I mean, and and and I I I think that so I
read that blog post.
I literally just emailed likefounders at warpstream.com.
Nice.
Uh and I was like, hey guys,like you seem to have some good
ideas.
You seem pretty smart.
Uh if you ever need a productperson, let me know.
And they replied being like,Yeah, yeah, we're we're three
people right now.
SPEAKER_01 (22:41):
Like that was a valid, that was a valid alias
they they had created.
SPEAKER_00 (22:44):
Yeah, too early.
No, no, I mean they said like inthe the CTA, like the call to
action at the bottom of the blogpost was like, you know, email
founders at if you'reinterested.
SPEAKER_01 (22:52):
Okay, okay, okay.
SPEAKER_00 (22:52):
Um, so I did.
I was like, I was interested fora different reason.
Uh just because it seems like avery interesting problem.
And it was also, you know, veryit was it was early, it was like
early in a in a startup's lifecycle.
It's just I wanted I'd neverdone that before.
I wanted to get you know thatexperience.
Um there was a lot going for it,basically, in my mind.
Um and yeah, so I I they theywere basically just yeah, yeah,
(23:16):
it's too early, like whatever,let's let's keep talking.
Uh and then like a week later,it was like, just kidding, we're
gonna raise a series A becauseVC's read this blog post and
they want to invest in ourcompany.
Would would you like to talkmore seriously?
I was like, Yeah, okay, cool.
Yeah, we'll do this.
Yeah.
Um yeah, so and and so I joinedI joined Warp Stream December
(23:36):
2023.
Um and yeah, about eight monthslater, we were, you know, in the
in the early uh early talks withConfluent and ended up joining
Confluent.
And the way that happened um waspretty interesting.
I mean it it at first, you know,that we didn't really know what
to make of it.
(23:56):
We're like, you know, what doesConfluent really want to do with
Warp Stream?
Like they've they've got freightclusters, like what do they what
do they want with us?
Um so we all we had a meeting inMountain View.
We all went to Mountain View andand met with um with you know a
bunch of people on the Confluentside.
And we kind of got someconfidence there that you know
the intent was not to like putus on a shelf, you know, that we
(24:18):
were a threat to their businessor anything like that.
Um we we felt like it was veryadditive and that like they they
you know the team very muchwanted like another line of
business.
The BYOC deployment model waslike the thing that made it
tick, and like that, you know,that that uh that was that was a
really good process for us to gothrough.
(24:40):
I think that if it was, youknow, if if if it was a
different situation with adifferent, you know, a different
uh you know parent company, Ithink that it would be uh
slightly slightly differentdynamic for sure.
Um, you know, and and you know,maybe maybe it would have worked
out differently or whatever, butwe didn't even you know think
that way.
We were just like, okay, Confinewants to add on this this new
(25:03):
line of business, we're a goodway to do that.
Yes.
Um and it was very additive, soit was it was good.
Yeah.
SPEAKER_01 (25:10):
Excellent.
How would you I love I love thatarc, I love that story.
If you had to summarize the likethe problem you solve, so kind
of warp stream is the answer.
What's the hardest problemyou're solved?
Well, warp stream.
You as the PM there, yeah, umsummarize that through that
(25:33):
lens, like as a as a as a as alife scale problem challenge,
cool thing that you're doing.
Um what is it that uh you do?
You you talked about being a PM,but like as a problem, what is
the problem?
SPEAKER_00 (25:52):
Do you mean like what what problem does
WarpStream solve as a product orsolving for okay?
SPEAKER_01 (25:58):
That's that's actually worth in case there's
anybody listening who doesn'tknow.
Sure.
Uh in fact, I'll just I'll I'lltake a swing at that.
You tell me you tell me.
Okay, go for it.
Yeah.
Uh WarpStream is a disklessKafka.
So um it is uh a system thatimplements the APIs and
semantics of Kafka, uh, butdoesn't store data on any
(26:20):
locally attached or networkattached storage.
Uh it's all stored in cloud blobstores.
This is a trend in datainfrastructure.
Pick a form of datainfrastructure, and you know, we
can identify the one or twogroups doing an open source
project to disclassify that.
Um and so it's a uh lower cost,higher latency way of being data
(26:43):
infrastructure that also givesyou BYOC characteristics.
So all of the data is stored in,say, your S3 bucket under your
cloud control, not under warpstreams, not under Confluence.
Like and I'll give me the warpstream pitch here.
Kind of cool thing about it isuh it's a very thorough BYOC
(27:04):
because like most BYOCimplementations, the vendor can
get in there and break glass ifthey have to and go in and fix
stuff in the data plane.
But here it's just there's noglass, it's it's it's out there
and nobody can do anything.
So yeah, it's BYOC DisclosKafka.
That's what that's what WarpStream does.
You're a PM there.
Um yeah, yeah.
So you've you've you've kind ofsolved that.
(27:28):
I guess I guess that's theproblem you've solved.
SPEAKER_00 (27:31):
Well, no, there there's a ton.
I mean, like, there's a ton ofpotential you know directions
that we could go.
There's infinite universes here.
Um but my my role here is tobasically help us pick the
direction that makes ourcustomers the most better off.
(27:54):
Um meaning, you know, what elsedo people want to use our
platform for?
The deployment model providessome interesting possibilities
for you know um the, you know,the like you were talking about.
It runs in your account, in yourVPC or in your data center, even
uh, and just egresses metadataout.
(28:14):
So with that in mind, likethere's some use cases where
maybe we could like lean intothat with some more product
features to make it um you knowgood for very highly sensitive
workloads that you know you youwouldn't you wouldn't ever want
to put on a cloud service.
So that's one sort of categoryof things we can do.
(28:35):
Um there's you know the the thescaling element of like the the
scalability of the of theplatform is another sort of
thing we can we can sort ofexplore.
For example, we recently sort ofrebuilt our storage engine to be
able to uh support um basicallyhuge amounts of data retention
(28:58):
because people there's a therewe had a couple of customers who
were saying basically that likewe have infinite retention on
compacted topics and we want tostore data forever.
We just expect this to keepgrowing forever.
Um and so you know, we need youto be able to support that.
And it's not it's not like okay,cool, now we can just like
forever into infinity supportall the data you could ever
(29:21):
store in a Kafka cluster, in awarp stream cluster, but we like
we 10x'd it.
We made it so that you couldstore you know huge amounts of
data.
Um, and you know, there'sthere's more work to be done
there to like figure out youknow, under the hood whether
that's like we we basically weextended the the timeline that
we have because as a streamingplatform, they're all always
(29:44):
writing more data in, and if itnever gets truncated, it just
keeps growing.
Um, you know, so there's like athere's some roadmap items
there.
I think the most exciting thingthough, right now that we're
working on uh is warp streamtable flow.
And so if if you're not familiarwith what table flow is, um
yeah, I mean this is an EA now.
SPEAKER_01 (30:03):
We've we've released it sort of an early access uh
we're we're speaking in um earlyto mid-November of 2025, if
you're correct listening tothis.
I don't know when this will bepublished.
Future timeline, yeah.
SPEAKER_00 (30:14):
Yeah, it might be GA by the time this this uh this
goes live.
SPEAKER_01 (30:17):
Uh probably will be GA by the time this goes live.
But yeah, long tail viewers inthe future.
SPEAKER_00 (30:21):
Uh yeah, yeah, yeah, exactly.
Uh so so yeah, so WarpStreamTableflow uh basically does the
same thing as confl and cloudtableflow.
It materializes your uh Kafkatopics as iceberg tables in
object storage.
Um Warpstream has the BYOCdeployment model, uh, which Tim
(30:45):
you described uh pretty well.
Um you know the data plane runsin your in the customer's VPC uh
with zero access delegated tous.
And so there's no cross accountI am privileges, so it all runs
in your VPC.
We don't have any access to theactual data.
Um the you know, your workloadegresses metadata to our control
(31:06):
plane, which is basically thebrain of the operation.
And so what we're doing withTableFlow uh is we basically
created a new cluster type thatum, you know, the the warp
stream agent, which is thereplacement for the broker, the
Kafka broker that runs in yourenvironment, uh instead of doing
the Kafka job, it does theiceberg job.
(31:27):
It writes parquet files, takesinput from a Kafka topic, writes
parquet files and objectstorage, builds the metadata
manifest thing, um, and shipsthe metadata about that back to
our control plane.
And just like Complent CloudTableflow, it's a fully managed
table, meaning like table flowhandles all of like the
(31:48):
tableflow agents handle all thebackground jobs, like uh file
compaction, compaction is thebiggest, you know, cleaning up
orphan files, deletes, thingslike that.
Um, and so it it's a it's it'smore of a managed service than,
for example, using like theKafka Iceberg connector, because
you still have to figure outlike how to manage the table and
(32:09):
how to manage the data in objectstorage and like how to do all
these operations, whichTableFull automates for you.
Yeah, um, so that's why we sayit's sort of a managed service
with a BYOC deployment model,um, because that you know it it
all runs in your in youraccount, but it automates a ton
of the tedious work that youneed to do if you if you were
trying to build an iceberg youknow data lake yourself.
SPEAKER_01 (32:30):
Uh no, I and you know I'm not here to to shill,
but table flow really is prettygreat.
It's it's a it's a it's a reallygood idea.
Um easy button.
SPEAKER_00 (32:40):
Yeah.
I mean the thing that peopleneed to understand about iceberg
is that it is a it's a tableformat, it's a building block.
Like it's not a data lake initself, it's a way to express
how to store data in objectstorage such that it can be used
as a data lake queryable with aSQL query engine.
Yes uh that's like that's notthat's like a that's like a uh a
(33:06):
small part of building adatabase.
Um what we've done, you know,Richie and Ryan like to say that
like we've table flow, warpstream tableflow is kind of the
bottom half of a database.
Uh it's like the back end likeprimitives, and we're gonna
build more on top of it, butlike that's um that's how to
think about it.
It's it's sort of uh it makesthe building blocks a lot more
(33:26):
useful and a lot more easy, likeeasy to uh to adopt.
SPEAKER_01 (33:30):
Right.
If you had to reflect, and it'searly.
I mean, warp stream is is youngand and growing and and and
still having its impact, but sofar, what do you think the
impact of it has been?
SPEAKER_00 (33:46):
Um Warpstream was if not the first uh one of the
first Diskless Kafka iterations.
Um and it I think it's not anexaggeration to say that that we
we pioneered um sort of thisidea.
(34:08):
We didn't know it at the timethat it was gonna be, you know,
that it was gonna have theeffect that it would have on the
industry, but we pioneered thisidea that you can sacrifice
latency in order to gain hugeamounts of cost savings,
specifically infrastructure costsavings.
Um and you know, now that yousort of poke your head up and
(34:29):
look around, there's there's youknow, there's an open source
Kafka, there's three open sourceKafka proposals currently in
debate.
There's um you know multiplevendors in the space.
Confluent has a managed servicethat uh you know basically is
built on a lot of the primitivesthat Wordstream is built on, um,
which is freight clusters andconfluent cloud.
(34:51):
It's like the fully hostedversion of what we do.
Um and so, you know, the the Ithink that the industry, I don't
know, I don't know if it waslike I don't know what the
direction of causality there is,um, but it definitely seems like
we influenced the direction ofthe industry by just taking the
first step to just be like,there's a possibility that
(35:13):
latency doesn't matter as muchas you think it does.
Let's just test that out andsee.
Um and I I think that you knowwe gained a lot of a lot of
certain traction um you knowearly on with that idea.
And I think that you know itcaught on very quickly.
Um it's been super interestingto be a part of it, yeah.
SPEAKER_01 (35:32):
My guest today has been Caleb Grillow.
Caleb, thanks for being a partof the Confluent Developer
Podcast.
SPEAKER_00 (35:38):
Thank you, Tim.
But we pioneered this idea that you can sacrifice
latency in order to gain hugeamounts of cost savings,
specifically infrastructure costsavings.
I mean, the thing that peopleneed to understand about iceberg
is that it is a it's a tableformat, it's a building block.
Everyone's like, oh, good PMsare right a lot.
You're you're wrong, like almostall the time.
You just have to be wrong in theright direction.
SPEAKER_01 (00:21):
Hey there, everybody, I'm Tim Berglund, and
welcome to Confluent Developer,the podcast where we explore the
fascinating journeys of softwaredevelopers tackling complex
problems.
In this episode, I talked toCaleb Grillo, the staff product
manager for Confluent WarpStream.
Caleb recounts his storystarting as a PM in the early
days of Confluent Cloud, talkingabout billing and a maturing
(00:42):
platform, and then later joiningan exciting young startup,
WarpStream, in late 2023.
WarpStream at the time was busyturning the world of Kafka
upside down, and Caleb tells thewhole story from the perspective
of a data-driven product managerwho is always wrong in
somebody's eyes.
Caleb tells a great story, andnow we get to listen.
(01:17):
Yeah.
Hey, tell us a little bit aboutyourself.
What uh what are you what areyou doing these days?
SPEAKER_00 (01:24):
Yeah, uh so I'm I'm Caleb.
Um I'm our uh product lead forthe Warp Stream product line at
Confluent, uh, which is our uhBYOC product line.
So bring your own cloud.
Um we uh we'll we'll get intoall the details of that, but
yeah, I I run our our productmanagement function um here.
(01:45):
And uh yeah, so that day to day,it's everything's different uh
every day.
Um so no no two days are reallythe same uh for me, but I work I
work across uh our engineeringteam to figure out you know what
we're what we're building andwhether it's the right thing for
our customers and make sure thatwe're prioritizing all the right
(02:06):
things.
Uh but I also do a fair amountof sort of analytics work, um
you know, digging into sort ofyou know how people are using
our product.
Um and I also work on sort ofthe business side, the GTM, um,
making sure that our sort ofcommercials make sense and and
sort of integration with thelarger confluent business and
(02:29):
all that.
So it's a pretty uh that's a lotknee knee deep, mile-wide type
role.
SPEAKER_01 (02:35):
What do you what do you mean analytics?
I thought being a PM was justvibes.
You're saying you use data?
SPEAKER_00 (02:39):
No, no, no, no.
Yeah, we uh so you know, all ofour consumption data and and all
of our usage data goes into umconfluence data warehouse, and
so we're able to sort of um dolike trend analysis on things
like revenue and you know costand make sure that our cost and
margin makes sense and all thatstuff.
SPEAKER_01 (02:58):
Yeah, so it's it's good it's good to have that
because there's I I've saidthere's probably a lot of
reasons why I'm not a productmanager, like you know, maybe
qualification or something, butI think you'd be a great one,
too.
SPEAKER_00 (03:09):
Well, here's the thing though.
SPEAKER_01 (03:10):
Here's the thing when you're a PM, like every
decision you make is to asignificant fraction of your
stakeholders, not just wrong,but like how could you be that
stupid wrong?
Yes.
So you're you're it's this thisprocess of being idiotically
wrong all the time.
In somebody's eye, like 40, 50percent of the people think that
(03:31):
all the time.
I got into a line of work whereI I get like the constant
affirmation that I need, likebeing on stage and stuff like
that.
So I think that's the it'sreally just an inner thing.
You know, I I need people totell me how awesome I am.
You need the strength ofcharacter to do the right thing,
even when they tell you you'rewrong.
SPEAKER_00 (03:47):
Oh you have that.
I think to be to be a reallygreat product person, you have
to be a glutton for punishment.
Yep.
Uh and you have to everyone'slike, oh, good PMs are right a
lot.
You're you're wrong, like almostall the time.
You just have to be wrong in theright direction.
SPEAKER_01 (04:02):
Yeah, and whether in terms of like objective results,
you're you're right or wrong,you you could come up with some
rubric for scoring that.
Somebody thinks you're wrong allthe time.
SPEAKER_00 (04:11):
Yeah, oh yeah, yeah, yeah.
SPEAKER_01 (04:12):
Yeah, you're just a guy, you're a moron all the
time.
And that I would I wouldstruggle.
Uh, what was what was your firstjob?
SPEAKER_00 (04:20):
Yeah, my first job out of out of college was not
um, well, my first job ever wasuh washing was washing windows
um in in high school.
Uh I had uh my one of mybrother's friends, I have an
older brother, one of mybrother's friends had a window
washing business and we'd goaround people's houses and wash
(04:40):
their windows.
That was my first job ever.
It was like a summer job.
Um my first serious job out ofcollege, uh, I worked for an
international development NGO.
Um and so yeah, it was reallyinteresting.
Um, we did a lot of work forlike USAID, the Agency for
International Development, uh,but also for sort of UN agencies
(05:03):
and other things like that.
Um and so, you know, and it wasreally global, like there was
there was there were projects inall kinds of different places
and different differentcountries and continents.
I didn't get to travel verymuch, uh, just you know, because
I I moved on to tech uh before Isort of got into travel-related
work.
But um that role was kind ofinteresting.
(05:24):
I was I was uh working onproposal budgeting.
So, you know, there's like atechnical portion of these, you
know, sort of project proposals,uh, and there's a cost
component.
And so I was working on the costcomponent, which is figuring
out, you know, how much does aToyota land cruiser cost in
Uganda, uh, you know, andputting that into budget
proposals and things like that.
(05:45):
Um so yeah, I mean it it was itwas sort of interesting.
I guess you could I I didn't atthe time think of it this way,
but you could draw someparallels between that and
product management, you know,making sure that all the pieces
fit together and make sense.
Um but uh yeah, I mean that thatwas my first job out of college.
Um yeah.
SPEAKER_01 (06:06):
There you go.
Window washing prepares you forproduct management only if your
customers berate you about whata terrible job you did washing
windows.
Yeah, that's right.
Um too slow.
Too slow.
You washed the wrong window panefirst, and uh the pattern you
used was wrong.
SPEAKER_00 (06:24):
I read in a book somewhere that you were supposed
to do it some other way.
SPEAKER_01 (06:27):
There you go.
This wiki this Wikipedia page.
Um tell us about I think youkind of hinted this, but what's
uh most interesting problem youever saw to date?
SPEAKER_00 (06:38):
Um yeah, so uh well I guess I can start with my my
uh my character arc um becausethat'll that'll sort of explain
my answer.
Uh love that, yeah.
I yeah, so so I I uh I got intoum tech after a couple years of
working for the InternationalDevelopment MBO NGO.
(06:59):
Um I I got into tech uh in sortof a business operations role
and then moved into productmanagement a couple years later.
Um and you know, one of my firstproduct jobs uh was um dealing
with a lot of the streaming dataat uh at an e-commerce site
(07:21):
that's not Amazon.
Uh and so um we were basicallybuilding a data platform for uh
for competitive intelligencedata.
So it was lots of high volumedata um that was being like
scraped from various places onthe internet, uh, basically
building a huge data lake of allthat information.
(07:41):
So teams like you know dataanalytics and data science could
figure out which prod productswere set were gonna sell well
against others and you know,things like that, uh, who which
sites were listing all thesedifferent products, how the
review counts were correlatingwith uh you know, and review
ratings were correlating withlike listing those those uh
(08:03):
products and surfacing them oncompetitor sites, basically
trying to learn what was goingon in the market from external
data sources instead of just ourown internal analytics.
And that uh that pipeline, likethe ingest pipeline, was Kafka.
Um so this was 2017, 2018.
SPEAKER_01 (08:20):
Um that was around the start of my my own Kafka
journey.
SPEAKER_00 (08:25):
Yeah, yeah, yeah, yeah.
Similar uh similar tracks, Iguess.
Um and so you know, I I came tolearn about Confluent when I was
in that role and we solved a lotof interesting problems.
Um, but I I kind of realizedlike, okay, there's like a more
scalable way to do this, like togo to Confluent and work on the
technology under underpinningthis this whole thing.
(08:46):
Um I just got super interestedin that.
And so I I ended up uh movingover to Confluent in May 2019.
Um and when I started, uh we hada sort of a you you'll remember
this, um, we had a sort ofdisjoint uh like product lineup.
There was in the there was acloud product that had gained a
lot of traction, but that cloudproduct was not really what you
(09:08):
would expect from likecommercially from a cloud
product.
Um, meaning like it wasn't likeAWS where you spin up an EC2
instance and you start beingbilled hourly for that instance
and you spin it down and it youstop being billed.
So there was there was like nopossibility of self-serve
anything uh in the in thatversion of our cloud product.
SPEAKER_01 (09:29):
It's it's it's in our defense back then, it's very
hard to make a cloud service.
And so, yeah, that was superprimitive.
Yeah.
SPEAKER_00 (09:37):
And it was it was like it was a proof of concept.
Like it was like uh, you know,will people use a streaming
platform hosted by somebodyelse?
Yes, that was proven uh thatthey would, and it kind of it
went too far.
SPEAKER_01 (09:51):
Uh we got carried away, went crazy.
SPEAKER_00 (09:55):
Yeah, yeah, it got carried away with the with the
proof of concept, and and thenit was time to make it a real
product.
Uh and so um when I came in atComplet, that was kind of the
state.
There was like, you know, thebig cloud business was this um
sort of interesting hybrid oflike a hosted service, um, but
it wasn't really what you'dexpect from like a fully managed
(10:17):
cloud product.
It was sort of just we'll runyour Kafka brokers for you.
SPEAKER_01 (10:21):
Um Kafka on some servers is yeah, yeah, yeah,
yeah, yeah.
SPEAKER_00 (10:26):
And we wanted like the vision for Comple and Cloud
was always to be way more thanthat.
Um servers.
And so yeah, and so so like thethe biggest the the the sort of
biggest blocker that we had,other than sort of the hard
technical things of you knowbuilding a a real cloud service,
um, we had the underpinnings ofthat, but like we were limited
(10:47):
by the commercial model.
Um and so that was sort of thethat was the big problem that
needed to be solved um back inyou know 2019 was like how do we
make the commercial model workum to in order to unlock sort of
the self-serve you know cloudvision of a cloud product that
(11:08):
everybody had.
Um and sort of central to thatwas building out the
consumption-based billingsystem, and so Confluent had a
consumption-based billing systemalready, but it was for their
self-serve cloud product.
So there were two differentproducts.
There was the non-self-serve,serious usage, dedicated
(11:31):
infrastructure hosted brokersproduct.
Yeah, you like it.
And there was another product,yeah, and there was another
product that was kind of justlike, you know, go sign up for
this thing, put in your creditcard, it's self-serve.
That was all multi-tenantclusters.
Um, so like there were all theunderpinnings that we needed to
sort of put everything together,but it was impossible to marry
those two worlds togetherwithout the commercial model
(11:54):
making sense.
SPEAKER_01 (11:54):
Okay.
SPEAKER_00 (11:55):
And so, in order to have a consumption-based
commercial model, you had tohave a consumption-based billing
system that could handle youknow all the different sort of
cluster types you'd need and allthe different infrastructure,
like there's you know, KafkaConnect, there's KSQL.
At the time, Flink wasn't there,but you need it to be sort of
able to support all thesedifferent um products and
(12:17):
different like cluster types,and you need to be able to meter
throughput and storage and likeall the different usage metrics.
Um, and so building out thefirst version of that was kind
of uh welcome to Confluent.
Here's here's this huge project,you need to go figure out how to
do it.
SPEAKER_01 (12:34):
Now, a quick word from our sponsor.
Confluent developer the podcastis brought to you by Confluent
Developer the website, which haseverything you need as a
developer of data streamingsystems.
And it's completely free.
We've got curriculum, hands-onexercises, tutorials, the online
data streaming engineercertification are also free.
(12:54):
A way to find a meetup near you,those are free, everything is
there.
I really want you to besuccessful in your journey as a
data streaming engineer, andthis is the site that has what
you need.
Check it out atdeveloper.confluent.io.
That's developer.confluent.io.
Now back to the show.
Um that was you and I werecoworkers then, but remind me,
(13:17):
were you a PM or an engineer?
What where were you then?
SPEAKER_00 (13:20):
I was a PM.
Yeah, yeah, yeah.
So I I I started I started aconfluent as a PM.
Uh and and so I yeah, that thatwas the first sort of thing to
figure out.
And that involved like it washugely cross-functional, uh,
which is a you know, hard thingin itself.
SPEAKER_01 (13:39):
Always.
SPEAKER_00 (13:39):
Meaning I had to work like with engineering, uh,
but also all the other PMs withrunning all their other
services, and also the salesteam and like to and the
marketing team, like the thewhole way things happened at
Confluent had to shift.
Uh, it wasn't just, oh, let'sjust build a billing system and
(13:59):
build it and they will come.
Like we had to make everythingwork.
The finance team, like we hadweekly meetings with like you
know, 10 different heads of X inthe room trying to figure out
sort of what we were doing.
Uh and so that was difficult.
It was also technicallydifficult because you know, like
(14:21):
any metering, like at the time,there weren't these nice SaaS
tools that you could buy to sortof solve the technical
underpinning of your billingsystem.
Like every company that wantedto do this had to build their
own system.
SPEAKER_01 (14:37):
Yeah.
SPEAKER_00 (14:37):
Right.
Yeah.
Uh and so we had already built,you know, before I got there,
there was there was like thissort of pre-existing metering
system.
Um, but it didn't support, forexample, having multiple cluster
tiers.
We couldn't have like adedicated cluster and a
multi-tenant cluster sittingnext to each other being built.
Uh, so we had to like build thatin.
(14:58):
Um there's also, you know, thisis getting into sort of boring
business stuff, but like there'suh there's a concept of like a
commit or like a minimumcommitment that you know a
customer would make in order tohave sort of a long-term deal
with Confluent.
Um and building that conceptinto the billing system.
(15:18):
Like we had to do that becauseyour usage gets metered, but you
can't just meter it at listprice if somebody's getting a
discount.
So we needed a concept of adiscount.
Like there were like all thesethings we had to model in and
sort of we were changing thewheels on the 747 as it was
coming in for a landing, youknow.
That's the metaphor.
SPEAKER_01 (15:37):
There's customers, customers using the platform at
this point.
SPEAKER_00 (15:40):
Yes, yeah, yeah, yeah, yeah.
That's true.
And then we had to come up witha strategy for migrating people
over and making sure that thepricing made sense, and like,
you know, we're just we'retrying to merge all these worlds
together.
Uh, so it's very, veryinteresting and very, very high
impact, uh, but also very hardum to get it right.
And you know, we didn't, I don'tthink we got it 100% right the
(16:01):
first time, uh, but we we builtit in a way that was like that
it could in theory support allthe future iterations that you
currently see in ConfluentCloud.
Yeah, um, you know, and it'stotally different now.
It's like a completely differentsystem.
Um, the technical side of it waswas also super interesting
because you know the theinfrastructure under the hood is
(16:24):
you know throwing off all sortsof metrics.
There, it's you know, you youcan measure all kinds of stuff.
Um, and figuring out like whatwhat pipelines, like what parts
of the observability pipelinewould be suitable for billing
was kind of a um, you know, thatwas a challenge in itself.
Uh figuring out which metricsmeant the things that they
(16:46):
needed to mean when it comes tobilling, because that's a
different problem thanobservability.
Yes.
Um, yes, I mean it it it was itwas uh it was a big undertaking,
but I I think that it, you know,I I learned a ton.
Uh it's a great, it's a greatintroduction to a company to be
like, hey, can we just likechange everything about it?
SPEAKER_01 (17:02):
Uh yeah.
Totally cross-functionally.
Different agendas, differentmotivation, different incentive
structures, different uhpersonalities, and and and you
make it all work.
SPEAKER_00 (17:14):
Yeah, yeah, yeah.
The technical side of it alsowas was super cool.
Uh, I don't want to minimizethat.
Like that that was um for mevery interesting to figure out
like how to translate all ofthese different requirements,
very specific, like you know uhlike gigabytes of cons of of
(17:36):
rights, like rights to thecluster like means something
specific.
There's like six differentmetrics that you could
potentially think would be youknow used for measuring rights
and what's the right one.
Uh and so you know, there'sthere's stuff like that.
And then you know, um different,like you know, I guess if uh
(17:57):
somebody wants to be able to doper second billing, it's like,
well, actually we want toaggregate things hourly and send
present it hourly to cut likethere there were there were lots
of arguments about like whatspecific things we did and and
we had to figure out like whatwas supportable with the current
platform and what when you havethings like you know bandwidth
limitations and things likethat, that implies uh a
(18:20):
measurement regime and a timegranular.
SPEAKER_01 (18:23):
There's all kinds of things that that have to be true
that are their own engineeringproblems to solve.
And dirty secret in any givensystem that has a thing like
that, you know, that might beover the whole day or something,
you know, that it isn'tnecessarily like an
instantaneous uh thing, it justdepends on what you've built.
(18:45):
And it's it's all its own way.
SPEAKER_00 (18:48):
Yeah, this was something that was interesting
that we always had to explain topeople.
It was like there's like aperiod of time where when you
like log into the UI and youlook at the billing screen,
there's a period of time wherethe amounts can change.
That like broke people's brainswhen you're trying to explain
this to customers.
They were like, and this was youknow, 2019, 2020, you know,
(19:09):
yeah, six years ago.
Uh people would would look atthat statement and be like,
wait, but I can I trust yourbilling system and be like,
yeah, yeah, we have all thesecorrectness checks, and
everything's like everything'sgood.
It's just that there can be somelate-arriving metrics.
Like this is a massivedistributed system of
distributed systems.
Like you can have stuff happen,but don't worry, like your
(19:30):
monthly bill is fine.
SPEAKER_01 (19:31):
Yeah, yeah.
It'll converge, it'll converge.
SPEAKER_00 (19:34):
Yeah, yeah, it'll converge, right?
SPEAKER_01 (19:37):
So you um you ended up fast-forwarding a little bit
at at Warp Stream.
Uh how'd you get there?
unknown (19:46):
Yeah.
SPEAKER_00 (19:46):
Yeah, so uh I had been at Confluent for about five
years.
Um and actually it it came aboutbasically because Rishi, uh the
one of the co-founders ofWarpstream, posted this blog
post, which is now very famous,uh, called Kafka is dead, long
live Kafka.
(20:07):
And all the ideas that they wereputting out there were just very
interesting.
They were basically saying thatlike there's this category of
workload where you know Kafka isa super low latency real-time
system.
There's this category ofworkload that doesn't need to be
super low latency, like it's notextremely latency sensitive.
(20:30):
And I remember this, you know,for years and years.
It was just like it wasn't evena thought that could cross your
mind that you coulddifferentiate a a product on
latency.
It was like, no, it's just fixedconstraint.
It's a real-time system, latencymust be small.
SPEAKER_01 (20:46):
It's um you know, is it is it zero yet?
Well, keep working, you know.
SPEAKER_00 (20:50):
Yeah, exactly.
And it's like it there, therewas never the insight of like
you can differentiate uh on thisdimension that like is a little
bit counterintuitive if you'rejust glancing at it.
But if you think about it for asecond, it's like if you're
running an observabilityplatform for customers, you
know, that's your product.
You your your customer doesn'tnotice if it takes a few hundred
(21:14):
milliseconds for your ingestpipeline to receive data.
Like it does not matter in thegrand scheme of things because
there's this whole pipeline thatneeds to happen that needs to
run to process that data beforeit can ever be displayed on a
graph in your product.
SPEAKER_01 (21:31):
And I I remember, I I think it was the summer of
2020.
Like you could check me on that,but reading that blog post and
just kind of warp stream comingon my radar and realizing
there's 23.
23.
Okay.
Okay.
Um realizing I don't know whatthe the like the demand
(21:51):
elasticity of uh of latencyreally is.
Like how many how many people ifif you could give them half the
price of Or a tenth of theprice, or you know, do that, but
it takes a second.
How many people care?
Like, I didn't know.
And it was just it was that samerevelation.
Like, wow, that could be 90% ofthe market that just doesn't
(22:13):
care.
I and and I still don't think wereally know that, but um it was
very interesting.
unknown (22:19):
Yeah.
SPEAKER_00 (22:20):
Yeah, I mean, and and and I I I think that so I
read that blog post.
I literally just emailed likefounders at warpstream.com.
Nice.
Uh and I was like, hey guys,like you seem to have some good
ideas.
You seem pretty smart.
Uh if you ever need a productperson, let me know.
And they replied being like,Yeah, yeah, we're we're three
people right now.
SPEAKER_01 (22:41):
Like that was a valid, that was a valid alias
they they had created.
SPEAKER_00 (22:44):
Yeah, too early.
No, no, I mean they said like inthe the CTA, like the call to
action at the bottom of the blogpost was like, you know, email
founders at if you'reinterested.
SPEAKER_01 (22:52):
Okay, okay, okay.
SPEAKER_00 (22:52):
Um, so I did.
I was like, I was interested fora different reason.
Uh just because it seems like avery interesting problem.
And it was also, you know, veryit was it was early, it was like
early in a in a startup's lifecycle.
It's just I wanted I'd neverdone that before.
I wanted to get you know thatexperience.
Um there was a lot going for it,basically, in my mind.
Um and yeah, so I I they theywere basically just yeah, yeah,
(23:16):
it's too early, like whatever,let's let's keep talking.
Uh and then like a week later,it was like, just kidding, we're
gonna raise a series A becauseVC's read this blog post and
they want to invest in ourcompany.
Would would you like to talkmore seriously?
I was like, Yeah, okay, cool.
Yeah, we'll do this.
Yeah.
Um yeah, so and and so I joinedI joined Warp Stream December
(23:36):
2023.
Um and yeah, about eight monthslater, we were, you know, in the
in the early uh early talks withConfluent and ended up joining
Confluent.
And the way that happened um waspretty interesting.
I mean it it at first, you know,that we didn't really know what
to make of it.
(23:56):
We're like, you know, what doesConfluent really want to do with
Warp Stream?
Like they've they've got freightclusters, like what do they what
do they want with us?
Um so we all we had a meeting inMountain View.
We all went to Mountain View andand met with um with you know a
bunch of people on the Confluentside.
And we kind of got someconfidence there that you know
the intent was not to like putus on a shelf, you know, that we
(24:18):
were a threat to their businessor anything like that.
Um we we felt like it was veryadditive and that like they they
you know the team very muchwanted like another line of
business.
The BYOC deployment model waslike the thing that made it
tick, and like that, you know,that that uh that was that was a
really good process for us to gothrough.
(24:40):
I think that if it was, youknow, if if if it was a
different situation with adifferent, you know, a different
uh you know parent company, Ithink that it would be uh
slightly slightly differentdynamic for sure.
Um, you know, and and you know,maybe maybe it would have worked
out differently or whatever, butwe didn't even you know think
that way.
We were just like, okay, Confinewants to add on this this new
(25:03):
line of business, we're a goodway to do that.
Yes.
Um and it was very additive, soit was it was good.
Yeah.
SPEAKER_01 (25:10):
Excellent.
How would you I love I love thatarc, I love that story.
If you had to summarize the likethe problem you solve, so kind
of warp stream is the answer.
What's the hardest problemyou're solved?
Well, warp stream.
You as the PM there, yeah, umsummarize that through that
(25:33):
lens, like as a as a as a as alife scale problem challenge,
cool thing that you're doing.
Um what is it that uh you do?
You you talked about being a PM,but like as a problem, what is
the problem?
SPEAKER_00 (25:52):
Do you mean like what what problem does
WarpStream solve as a product orsolving for okay?
SPEAKER_01 (25:58):
That's that's actually worth in case there's
anybody listening who doesn'tknow.
Sure.
Uh in fact, I'll just I'll I'lltake a swing at that.
You tell me you tell me.
Okay, go for it.
Yeah.
Uh WarpStream is a disklessKafka.
So um it is uh a system thatimplements the APIs and
semantics of Kafka, uh, butdoesn't store data on any
(26:20):
locally attached or networkattached storage.
Uh it's all stored in cloud blobstores.
This is a trend in datainfrastructure.
Pick a form of datainfrastructure, and you know, we
can identify the one or twogroups doing an open source
project to disclassify that.
Um and so it's a uh lower cost,higher latency way of being data
(26:43):
infrastructure that also givesyou BYOC characteristics.
So all of the data is stored in,say, your S3 bucket under your
cloud control, not under warpstreams, not under Confluence.
Like and I'll give me the warpstream pitch here.
Kind of cool thing about it isuh it's a very thorough BYOC
(27:04):
because like most BYOCimplementations, the vendor can
get in there and break glass ifthey have to and go in and fix
stuff in the data plane.
But here it's just there's noglass, it's it's it's out there
and nobody can do anything.
So yeah, it's BYOC DisclosKafka.
That's what that's what WarpStream does.
You're a PM there.
Um yeah, yeah.
So you've you've you've kind ofsolved that.
(27:28):
I guess I guess that's theproblem you've solved.
SPEAKER_00 (27:31):
Well, no, there there's a ton.
I mean, like, there's a ton ofpotential you know directions
that we could go.
There's infinite universes here.
Um but my my role here is tobasically help us pick the
direction that makes ourcustomers the most better off.
(27:54):
Um meaning, you know, what elsedo people want to use our
platform for?
The deployment model providessome interesting possibilities
for you know um the, you know,the like you were talking about.
It runs in your account, in yourVPC or in your data center, even
uh, and just egresses metadataout.
(28:14):
So with that in mind, likethere's some use cases where
maybe we could like lean intothat with some more product
features to make it um you knowgood for very highly sensitive
workloads that you know you youwouldn't you wouldn't ever want
to put on a cloud service.
So that's one sort of categoryof things we can do.
(28:35):
Um there's you know the the thescaling element of like the the
scalability of the of theplatform is another sort of
thing we can we can sort ofexplore.
For example, we recently sort ofrebuilt our storage engine to be
able to uh support um basicallyhuge amounts of data retention
(28:58):
because people there's a therewe had a couple of customers who
were saying basically that likewe have infinite retention on
compacted topics and we want tostore data forever.
We just expect this to keepgrowing forever.
Um and so you know, we need youto be able to support that.
And it's not it's not like okay,cool, now we can just like
forever into infinity supportall the data you could ever
(29:21):
store in a Kafka cluster, in awarp stream cluster, but we like
we 10x'd it.
We made it so that you couldstore you know huge amounts of
data.
Um, and you know, there'sthere's more work to be done
there to like figure out youknow, under the hood whether
that's like we we basically weextended the the timeline that
we have because as a streamingplatform, they're all always
(29:44):
writing more data in, and if itnever gets truncated, it just
keeps growing.
Um, you know, so there's like athere's some roadmap items
there.
I think the most exciting thingthough, right now that we're
working on uh is warp streamtable flow.
And so if if you're not familiarwith what table flow is, um
yeah, I mean this is an EA now.
SPEAKER_01 (30:03):
We've we've released it sort of an early access uh
we're we're speaking in um earlyto mid-November of 2025, if
you're correct listening tothis.
I don't know when this will bepublished.
Future timeline, yeah.
SPEAKER_00 (30:14):
Yeah, it might be GA by the time this this uh this
goes live.
SPEAKER_01 (30:17):
Uh probably will be GA by the time this goes live.
But yeah, long tail viewers inthe future.
SPEAKER_00 (30:21):
Uh yeah, yeah, yeah, exactly.
Uh so so yeah, so WarpStreamTableflow uh basically does the
same thing as confl and cloudtableflow.
It materializes your uh Kafkatopics as iceberg tables in
object storage.
Um Warpstream has the BYOCdeployment model, uh, which Tim
(30:45):
you described uh pretty well.
Um you know the data plane runsin your in the customer's VPC uh
with zero access delegated tous.
And so there's no cross accountI am privileges, so it all runs
in your VPC.
We don't have any access to theactual data.
Um the you know, your workloadegresses metadata to our control
(31:06):
plane, which is basically thebrain of the operation.
And so what we're doing withTableFlow uh is we basically
created a new cluster type thatum, you know, the the warp
stream agent, which is thereplacement for the broker, the
Kafka broker that runs in yourenvironment, uh instead of doing
the Kafka job, it does theiceberg job.
(31:27):
It writes parquet files, takesinput from a Kafka topic, writes
parquet files and objectstorage, builds the metadata
manifest thing, um, and shipsthe metadata about that back to
our control plane.
And just like Complent CloudTableflow, it's a fully managed
table, meaning like table flowhandles all of like the
(31:48):
tableflow agents handle all thebackground jobs, like uh file
compaction, compaction is thebiggest, you know, cleaning up
orphan files, deletes, thingslike that.
Um, and so it it's a it's it'smore of a managed service than,
for example, using like theKafka Iceberg connector, because
you still have to figure outlike how to manage the table and
(32:09):
how to manage the data in objectstorage and like how to do all
these operations, whichTableFull automates for you.
Yeah, um, so that's why we sayit's sort of a managed service
with a BYOC deployment model,um, because that you know it it
all runs in your in youraccount, but it automates a ton
of the tedious work that youneed to do if you if you were
trying to build an iceberg youknow data lake yourself.
SPEAKER_01 (32:30):
Uh no, I and you know I'm not here to to shill,
but table flow really is prettygreat.
It's it's a it's a it's a reallygood idea.
Um easy button.
SPEAKER_00 (32:40):
Yeah.
I mean the thing that peopleneed to understand about iceberg
is that it is a it's a tableformat, it's a building block.
Like it's not a data lake initself, it's a way to express
how to store data in objectstorage such that it can be used
as a data lake queryable with aSQL query engine.
Yes uh that's like that's notthat's like a that's like a uh a
(33:06):
small part of building adatabase.
Um what we've done, you know,Richie and Ryan like to say that
like we've table flow, warpstream tableflow is kind of the
bottom half of a database.
Uh it's like the back end likeprimitives, and we're gonna
build more on top of it, butlike that's um that's how to
think about it.
It's it's sort of uh it makesthe building blocks a lot more
(33:26):
useful and a lot more easy, likeeasy to uh to adopt.
SPEAKER_01 (33:30):
Right.
If you had to reflect, and it'searly.
I mean, warp stream is is youngand and growing and and and
still having its impact, but sofar, what do you think the
impact of it has been?
SPEAKER_00 (33:46):
Um Warpstream was if not the first uh one of the
first Diskless Kafka iterations.
Um and it I think it's not anexaggeration to say that that we
we pioneered um sort of thisidea.
(34:08):
We didn't know it at the timethat it was gonna be, you know,
that it was gonna have theeffect that it would have on the
industry, but we pioneered thisidea that you can sacrifice
latency in order to gain hugeamounts of cost savings,
specifically infrastructure costsavings.
Um and you know, now that yousort of poke your head up and
(34:29):
look around, there's there's youknow, there's an open source
Kafka, there's three open sourceKafka proposals currently in
debate.
There's um you know multiplevendors in the space.
Confluent has a managed servicethat uh you know basically is
built on a lot of the primitivesthat Wordstream is built on, um,
which is freight clusters andconfluent cloud.
(34:51):
It's like the fully hostedversion of what we do.
Um and so, you know, the the Ithink that the industry, I don't
know, I don't know if it waslike I don't know what the
direction of causality there is,um, but it definitely seems like
we influenced the direction ofthe industry by just taking the
first step to just be like,there's a possibility that
(35:13):
latency doesn't matter as muchas you think it does.
Let's just test that out andsee.
Um and I I think that you knowwe gained a lot of a lot of
certain traction um you knowearly on with that idea.
And I think that you know itcaught on very quickly.
Um it's been super interestingto be a part of it, yeah.
SPEAKER_01 (35:32):
My guest today has been Caleb Grillow.
Caleb, thanks for being a partof the Confluent Developer
Podcast.
SPEAKER_00 (35:38):
Thank you, Tim.
Popular Podcasts
Betrayal Weekly
Betrayal Weekly is back for a new season. Every Thursday, Betrayal Weekly shares first-hand accounts of broken trust, shocking deceptions, and the trail of destruction they leave behind. Hosted by Andrea Gunning, this weekly ongoing series digs into real-life stories of betrayal and the aftermath. From stories of double lives to dark discoveries, these are cautionary tales and accounts of resilience against all odds. From the producers of the critically acclaimed Betrayal series, Betrayal Weekly drops new episodes every Thursday. If you would like to share your story, you can reach out to the Betrayal Team by emailing them at betrayalpod@gmail.com and follow us on Instagram at @betrayalpod and @glasspodcasts. Please join our Substack for additional exclusive content, curated book recommendations, and community discussions. Sign up FREE by clicking this link Beyond Betrayal Substack. Join our community dedicated to truth, resilience, and healing. Your voice matters! Be a part of our Betrayal journey on Substack.
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com