00:00hi and welcome to the a 16z podcast
00:02today's episode based on a panel by in
00:05for developers at our recent summit
00:07event dives into the details of what
00:09happens when software eats development
00:11and how the entire development lifecycle
00:12is adapting from how you design your
00:15software how you build it how you
00:16release it how you configure it to how
00:18you monitor it moderated by martine
00:21casado the conversation includes in the
00:23order in which she'll hear their voices
00:25Florian Liebert CEO and co-founder of
00:27mesosphere Matt Billman CEO and
00:30co-founder of net liffe i and karthik
00:32rao cofounder and CEO of signal FX this
00:36is for me one of the most interesting
00:37questions and trends in all of IT buying
00:39which it used to be the case have you
00:40sold something that was a piece of
00:42infrastructure you sell it to the ops
00:43person or a core IT or whatever more and
00:46more developers are involved in this
00:48purchasing that's in and so this kind of
00:49theory is you can think of it you can be
00:50like okay there's core IT that's buying
00:52there's this new group maybe that's kind
00:53of like a devil opposite group or
00:55basically in the end it's the developers
00:56that do all the buying and have all the
00:58control flow I'd love for you to talk a
01:01little bit about how you seen this
01:02evolution I mean with measure you guys
01:03have been in the thick of this is it
01:04core IT is it developers is it something
01:06in between so our software is the the
01:09end users of the software actually
01:11operators and also data scientists so
01:14those are the folks that actually use
01:16our product day-to-day but then of
01:18course they install these platform
01:19services and platform services are for
01:21example distributed databases message
01:24queues and many other things that you
01:26find on Amazon that you find on Google
01:27cloud and what are these things used for
01:29well they're four basic pillars
01:31specially and one is transporting data
01:33the other one is processing data the
01:36next one is storing data and then the
01:38other one is serving data backup and you
01:40have basically many implementations
01:42within each of these pillars and of
01:44course again the developers are the
01:46end-users that take some of these
01:47components in order to assemble the
01:49applications but of course also data
01:51scientists will install notebooks
01:54they'll install spark they'll install
01:55Hadoop and then use it and the operator
01:58of the entire platform is usually the
02:00operator of the DevOps person so I want
02:02to make sure I understand the layered
02:03cake so you're basically saying okay
02:04that was core IT that's kind of
02:06antiquated DevOps is eating core I T
02:08yeah developers reading DevOps and now
02:10you're just said that data scientists
02:11are eating developers is that the
02:13lack of power actually I think right in
02:15the future hopefully we'll be able to
02:17click together applications you'll have
02:18more and more building blocks more and
02:20more applications and frankly it's more
02:22joyful to write Visual Basic than it is
02:24to write go right there you actually
02:26achieve business results I mean the
02:28seriously right like getting a spark job
02:30you see the output right away whereas if
02:32you write a large go application I mean
02:34it's a lot of work and you're just
02:36dealing with plumbing so today when you
02:38build a mobile app you build everything
02:39in the mobile app and you run it on your
02:41phone and then for like core services
02:42it'll call up to api's or micro services
02:44on the backend like 1200 Stryper or
02:46whatever so now the fire is doing this
02:47for the web so basically you create a
02:49web page that's a thick front-end and
02:50then you call micro services in the
02:52backend so you can you know follow the
02:54general model of that so this is kind of
02:56like developers eat web development so
02:58how do you view the evolution of this is
03:00the buyer changing is the influencer
03:02changing how is that ecosystem evolving
03:04so one of the things is again like the
03:06developer themself changing we used to
03:08have this very clear distinction between
03:09the back-end developers and then
03:11front-end developers that work with the
03:13web port barely release developers they
03:15would mostly take some design file and
03:17then cut it up and hand it over to a
03:20back-end developer that would integrate
03:21that into some big monolithic
03:24application and now we're seeing more of
03:26that almost this show basically in the
03:28form of JavaScript where we didn't have
03:29this whole new like world front-end
03:31developers that become the actual web
03:33developers and through latch degrees
03:35that just gluing together a lot of
03:38different micro services that are
03:40already ready on the backend instead of
03:42like building these flash monoliths okay
03:45I wanna make sure that I understand this
03:46deck so front-end developers are they
03:49obviating designers or they're still is
03:51it just basically now you have designers
03:53you got front-end developers then you've
03:54got people do micro service and the
03:55back-end developers are becoming micro
03:57service up is that how you view the yeah
03:59and in many of these cases the need for
04:02begging developers to sort of shrinking
04:03this kind of going away to one agency
04:08that's really been a pioneer of adopting
04:10this model you talked about off building
04:12sites the juice to have like 55
04:14developers and steps where
04:1635 of those were back-end developers and
04:18now they have around 30 free developers
04:21and two of them are begging developers
04:22from all the rest just works in this
04:24friend and layer so that's where the big
04:26shift has happened bitten
04:27just like we suddenly had a whole world
04:29of app developers merging with the
04:31iPhone now we have this whole new world
04:33of front-end developers that are become
04:35gets to web developer Karthik so you
04:37don't have very similar backgrounds both
04:38worked at VMware or what kind of core
04:40ops and then you're very much focused on
04:42ops for what you're doing but it seems
04:44to me that the abstractions have evolved
04:46like what's important to highlight I
04:48mean is that the case as developers
04:49become more in the picture as API is
04:51make sure it's good more in the pictures
04:52the type of information unit service the
04:54ops team evolving or is it still a kind
04:56of IP addresses and low-level stuff now
04:58I think it's higher level you know
04:59developers are more and more involved
05:02and in fact in new applications are
05:04making the technology decisions for the
05:06core runtime by you know if you look at
05:08organizations kind of at maturity it
05:10doesn't make sense to have 10 or 20
05:13different teams all running isolated
05:16kafka clusters and relearning best
05:18practices around how to scale and make
05:20it resilient and so it a lot of these
05:22organizations you do find a centralized
05:24team that will develop best practices
05:26around some of these core infrastructure
05:27services you know if you're running in a
05:29cloud you can maybe just use an API to
05:31leverage some of these core services
05:32from an Amazon or Google but within an
05:35enterprise oftentimes you'll still find
05:37these functions it's just they're
05:38focusing on a different kind of
05:39technology than you know the traditional
05:41enterprise applications and if you get
05:44on this area what do you think is a core
05:45bit of insight people look for in this
05:46kind of new era of micro services well I
05:49think the big thing that's changed is
05:51the velocity of application development
05:52is just at an altogether different level
05:54right instead of doing two or three
05:56updates a year you could be doing
05:57thousands or tens of thousands of
05:59updates a year and so you know the
06:01entire development lifecycle deployment
06:03lifecycle the entire toolset is evolving
06:06to support velocity so how you design
06:08your software how you build it how you
06:10release it how you configure it how you
06:11monitor it's all adapting so in our
06:13world what people are looking for is to
06:16get insight into what's changing in the
06:18environment because there's so much
06:19change happening you want to be a lot
06:21more proactive and identifying a
06:22destructive change that you can roll it
06:24back very quickly and so having the
06:26analytics around all of your telemetry
06:27data to practically identify and surface
06:29those trends becomes absolutely critical
06:31so that's the key insight that our
06:32custom pasa typically looking for so I'm
06:34glad you brought the developer lifecycle
06:36so something is actually unique across
06:37all of these companies is they integrate
06:39with a developer lifecycle so just as
06:41set up very quickly then I'd love for
06:42you guys to comment about how this is
06:44kind of changed where you think about
06:45business developers now follow a pretty
06:47common recipe for developing it's called
06:49the CI CD pipeline we have continuous
06:51integration continuous development and
06:53they use kind of pretty standard set of
06:54tools and more and more we see startups
06:56coming in that are traditional
06:57infrastructure startups which 10 years
06:59ago they'd sell a box like they sell a
07:00router or something like that or a piece
07:02of software and these days they
07:03integrate in the CI DC pipeline and
07:05they're talking to developers it's a
07:07massive massive shift of Enterprise
07:08buying and it's also interesting because
07:10developers if you're part of the code or
07:12you're close to the code you have a lot
07:13more information and semantics so I know
07:15your component is directly into get so
07:18how do you think about the CI CD
07:19pipeline a and how you build that LIF i
07:21but also like how this is gonna
07:22influence infrastructure companies going
07:23forward it has been such a massive trend
07:26that's been adopted by developers
07:29universally and that's become not just
07:31their version control system but their
07:33main workflow Indian and the main way of
07:35collaboration and communication even in
07:38issues around GUID and so on so what
07:41we're doing is really taking the way
07:42they work there and then extrapolating
07:45it all the way out through the
07:46deployment and infrastructure to really
07:47mean that developers working with metal
07:49if I essentially just work and get when
07:51they make a new push and pull request to
07:53get we create a new staging environment
07:55for them and give them a new URL where
07:56they can view that right when they merge
07:58that into master we deployed that
08:00version of the site and it goes live
08:01when they change settings for caching
08:04they do it just by writing a file in the
08:06git repository and push it when they
08:08work with content management they do it
08:10just by putting a UI on top of the git
08:13repository and centrally and giving
08:15normal people that way you to edit the
08:17content and contribute to the same like
08:19workflow so we're really just thinking
08:21about how can we take that really
08:23efficient workflow that developers are
08:25working with and then just make
08:26everything happen from there so in net
08:29liffe I if somebody is editing like a
08:31developers making a website they make
08:33the website as soon as they push it in
08:34to get it get slurped up identify thrown
08:37in a CDN and distributed across the
08:39world and that does the actual
08:40operations and deployment do you think
08:42the separation between developers and
08:43deployment operations goes away do you
08:45think like this is how things go in the
08:47future or is there still another team
08:48that's doing the operate I mean it kind
08:51of dangerous to me ah yeah so
08:53the trick is to figure out ways to make
08:55it really safe you can see every pull
08:57request in a branch so you know what's
08:59going live before you take it live in
09:01edify everything is a mutable deploy you
09:03can roll back in an instant and we sort
09:05of build a whole UI around that
09:07management of the deployment and that in
09:09a certain degree replaces all the
09:10existing like release engineers and so
09:12on for that whole part of the stack and
09:15I think we'll see that more and more
09:17this whole notion that if you adopt a
09:19certain workflow then a system can kick
09:22in and automate the whole process around
09:24it so developers are now eating kind of
09:26deployment engine operators I love I
09:28mean again you you actually seen the
09:29full evolution of this over yesterday ok
09:31like how do you view the CIC be
09:32affecting how you do business or people
09:34run things or ships in the industry it's
09:36an interesting question so I used to
09:38work at Airbnb and everybody search was
09:40actually really difficult search for
09:42twitter is really easy you always show
09:43everybody pretty much the same results
09:45but for Airbnb if you do that then if
09:47the first person sees a result for San
09:50Francisco clicks on the booking the
09:52second person does the same then the
09:54booking might be gone so if this happens
09:56multiple times the users actually will
09:59so what we had to do with Airbnb every
10:01time we rolled out a new version of
10:02search we had to do a/b testing and we
10:04had to make sure that actually then
10:06that's a big part actually of CI cdza
10:08push out automatically a new version and
10:10then you redirect 5% of the traffic
10:12towards this new version and then you
10:15actually scale that version up to 10%
10:16scale the other one back and so forth
10:18and when you're running tens or hundreds
10:22of experiments like that a day
10:24automation is really key it's really key
10:26to actually automate every part it's
10:29really key to actually have the metrics
10:30in place that if something goes wrong
10:32that you can use tools like yours in
10:34order to do automatic rollback but in
10:36general you want to automate as much as
10:38possible because humans are the biggest
10:40source of errors and respect this kind
10:42of this thing I've been pushing on which
10:44is I just kind of so cares what the
10:45future looks like it's it literally just
10:47devs of different flavors like this is a
10:49dev that does front-end this is a dev
10:50that does automation but they're all
10:52software developers and part of the same
10:54CI CD pipeline or is it still gonna be
10:57you know your operations your expertise
10:58is running somebody else's software
11:00versus developing the software so I mean
11:02I think the whole operator is going to
11:04go away over time like with cloud with
11:07tools like do you think so you think
11:08operations yeah I think I think it's
11:09gonna things like we're gonna go towards
11:11the world where a lot of that stuff is
11:12autonomous that's trillion dollar
11:14business you know just I mean I actually
11:16agree with you I just want you know like
11:18to actually take a moment and understand
11:20so I mean I give you an example I mean
11:22in the past when you were set up as
11:23sharted well yeah shard of database
11:25structure we had DBA is actually making
11:28sure that when one of them goes down
11:29they actually do certain procedures and
11:32restore it somewhere else it was a lot
11:34of manual work and usually whenever you
11:36had humans involved the Twitter things
11:37went wrong that's way also the fail
11:39whale a lot there were too many people
11:40involved but now when you think about it
11:42when you have things like RDS or you
11:45have Cassandra the runs of DCs sorry for
11:48the shameless plug but if you have
11:50systems where a lot of the recovery of
11:54failures is actually built in and the
11:56level of automation is really high you
11:58actually don't need these DBA anymore
11:59sitting around and screwing stuff up I
12:01still think there's an element of
12:03operations around these workloads though
12:04right it's just the skillset and the
12:06type of people who are doing it are
12:07different it's not your classic DBA
12:09sysadmin it's gonna be someone more of
12:11an engineering skill set is this
12:13somebody that understands code and api's
12:15and whatever is that someone that
12:17understands running application like for
12:18me if a line is is do you write your own
12:20application is your expertise primarily
12:22writing an application or running
12:24somebody else's application I mean well
12:25let's take Cassandra as an example
12:27internally we've got a large Cassandra
12:29cluster and the developer who designed
12:30it for our particular workload is one of
12:32our most senior best engineers I don't
12:34want him doing operations on our
12:37Cassandra cluster so we've got a
12:38different skill set they're not quite
12:40the same level of seniority as appalled
12:42by Wright yeah see you should use our
12:44software are you taking advantage
12:49standardization the C ICD pipeline for
12:52for signal effects it seems like a very
12:54obvious way to kind of get access to
12:56signals yes C ICD and containers have
12:59completely changed the entire downstream
13:02systems management process you know you
13:04think about with containers the
13:05developer can now deploy directly into
13:07production environments you have a high
13:08rate of change direct deployment
13:10especially the mutable infrastructure
13:11you're seeing people effectively replace
13:13an entire application with a new set of
13:15containers and then you know
13:16decommissioning the old set so there's a
13:18massive amount of change and there's a
13:20turn in the environment and so being
13:23able to collect the signal so that you
13:25can more you know proactively identify
13:28what's different so a lot of our
13:29customers are doing not quite a be
13:31testing but canary deployments Bluegreen
13:33deployments where they're looking at
13:34metrics for the new set of
13:36infrastructure the new application
13:37version versus the old identifying
13:40outliers anomalies and then rolling back
13:41very quickly and do you require code
13:43instrumentation to do that for the
13:45higher-level application metrics the
13:47custom application metrics yes most of
13:49the custom it seems to me like another
13:51line that's five for kidding the
13:53industry here is the old model the
13:55interface is like IP or like the process
13:58or the file system or the database but
14:00now because see ICD is becoming
14:02standardized the interface is actually
14:03the programming language which is so
14:05powerful it used to be the case of I was
14:07gonna provide something for signaling
14:08for example I couldn't be part of the
14:10code so I'd either sit on the wire and
14:12I'd use network signals or maybe I'd use
14:13the file system or maybe some process
14:15performance characteristics but you can
14:17actually integrate in the code because
14:18the ICD and linkages are becoming
14:20standard so the zoom allows you to
14:21provide better visibility than you can
14:22before so playing this out does every
14:24box become basically a library and
14:26somebody's code or they're basically
14:27like the old set of vendors that we're
14:29providing boxes and this and that do
14:30they just basically become software
14:31libraries yes the richness of data that
14:33we can collect now is substantially
14:34better just given that the developer
14:36does have an incentive to instrument his
14:38or her own code because if something
14:40goes wrong they're the ones that are
14:41gonna get paged so bad do you see that
14:43you're appealing to a subset of the
14:46community that's growing or is it more
14:48its everybody and they're just learning
14:50are we seeing an emergence of a new kind
14:51of user a new buyer that understands
14:53development that understands programming
14:55or is everybody slowly moving towards it
14:57because they're gonna be older than
14:58yours everybody can become the new I
15:00think it's very much moving in
15:02concentric circles from a group of early
15:04adopters typically in the B area around
15:07open-source projects and so I'm that's
15:09been some of our huge adopter set in
15:11Italy if I write and then sort of
15:12spreads out deeper and deeper into the
15:14whole ecosystem where of course like you
15:17always have these curves of like
15:19adoptions that takes very very long to
15:21trail off so at this point 20% of all
15:24websites are still built with WordPress
15:26and everybody kind of knows it's a bad
15:28idea but it's even still slowly growing
15:31and I think we're about to see a pretty
15:34where the architecture of that will
15:36shift and that I think will affect like
15:38basically all web developers over the
15:40next ten years okay so for those that
15:43are running large companies that are
15:44moving towards containers into micro
15:45services the other understand that see
15:47ICD is evolving in the entire
15:49infrastructure is evolving to support
15:50that like is there any guidance while
15:52you would provide to someone running a
15:55business especially a legacy business
15:56and how to get ready for this and how
15:57they adopted what to experiment what to
15:59look into yes I don't think there's
16:01going to be this rip and replace so
16:03we're dealing with a lot of financial
16:04services companies and if you log on to
16:06your bank account or if you do an ATM
16:08transaction that still runs on a
16:10mainframe and that's probably not going
16:11to go away because I think a lot of the
16:13tools that have emerged they don't have
16:15the same level of consistency that the
16:17mainframe has but I think what's going
16:19to happen over time is as a bank for
16:21example rolls out new applications and
16:23those applications are client facing and
16:25they have more traffic you can introduce
16:28caching layers that actually do a read
16:30from the mainframe and then they can
16:32serve many many more clients so I think
16:34we're not going to see a rip and replace
16:36I think we're going to see more and more
16:38systems build around the old systems and
16:40for some of the core infrastructure I
16:42think even in 20 or 30 years we'll still
16:44have mainframes I agree that a lot of
16:46this will happen as gradual transitions
16:48right and out of the hole advantage of
16:50this micro service world is that you can
16:52actually split things into smaller
16:54services from that also means you can
16:55start building out some kind of find
16:57layer that now consumes your old
17:01mainframe stuff wrapped in a couple of
17:03smaller services but I think we'll see a
17:04lot of those gradual transition paths
17:06and old stuff sticks around for a very
17:09long time one piece of feedback I would
17:11have based on everything we've seen with
17:13enterprises is this whole model of CI CD
17:16micro services containers it's all
17:17around optimizing for speed and you know
17:20high velocity and speed and what a lot
17:23of people don't realize is the entire
17:25toolset has to evolve to enable you to
17:28catch issues more quickly and to roll
17:30them back lis as quickly as possible so
17:32that's the fundamental philosophical
17:33change in the toolset and you know from
17:35the very beginning and how you develop
17:37your code all the way through how you
17:38support manage it that's we're seeing a
17:41lot of interesting new technologies that
17:42are supporting this this very different
17:44model of software delivery so I you know
17:46urge everyone to take a look at the
17:48spectrum from build all the way to
17:50release manage and support and how
17:52that's going to change in this new world
17:53yeah please help me thank the panelists