Salut tout le monde, bienvenue à la fête de la fête de la podcast.
Google est un podcast sur la compétition de l'engineering et de la production de la

Je suis votre host, Steve McGee.
Cette fête est de nos amis et de nos taux de la France.
C'est tout pour ce qui est venu dans le space de la SRE, de la nouvelle technologie, de
les processus modernisés.
Et bien sûr, la partie la plus importante est la fête que nous avons faite.
Alors, bonsoir à tous et à vous de vous rappeler, j'espère que ce n'est pas une stratégie.
Salut, bienvenue à tous et à la fête de la podcast.
C'est un podcast sur la production de la SRE.
C'est un bon point.
On a un guest qui était de la façon de la suite.
Adam, vous avez aidé à commencer la fête quand c'était strictement interne.
Je pense que c'était vous et John, je vous rappelle.
C'est vrai.
C'est vrai.
Et puis, on a aussi...
Matt est ici.
Bonjour, Matt.
Comment ça va?
Bonjour.
Et je suis Steve, je dois dire que je n'ai pas de nom aussi.
Adam et moi avons rencontré dans l'office de la SRE à Google, en retour du lieu de l'olden.
Mais je pense que vous avez travaillé à d'autres places avant ça et travaillé à différents
places depuis que c'est vrai.
C'est vrai.

C'est le rohmer et aussi la vérité.
Oui, je suis à Google pour longtemps.
Je suis venu par la acquisition de l'ITA Software.
Je suis en train de mettre mon ITA Software sur le point, ce qui est confusant.
C'est le plane, c'est le plane.
Oui, le plane.

Oui, on a fait une search de flight et on a fait un réservation system.
Je suis sur le côté de l'osage de la réservation.
Google a bâti nous, surtout pour notre flight à l'autre côté, mais on a appris les flights
de Google à cet endroit.
Je suis au bureau de la campagne et j'ai rentré à San Francisco où j'ai été avec
Steve.
C'est plus de 10 ans, plus que ça?
12 ans ou 13 ans?
Je ne vais pas tenter de le faire.
C'est beaucoup de temps.
C'est beaucoup de temps.
Bien sûr, oui, envers quand nous étions tous les deux, on était tous les deux.
Et puis, je suis venu par Google pour longtemps, j'ai aimé travailler avec des gens comme
Steve et Paul, qui ne les connaissent pas, mais ils sont ici.
Et tous les gens, c'est un groupe superbe.
C'est un de mes company, j'ai aimé commencer et construire des start-ups.
C'est cool.
Adam a donné un très bon talk, je pense, sur Internet, qui est sur le point de voir
comment les searchés de flight sont compliqués, même si ça ne se sent pas comme ça.
Je vous recommande de le vérifier si vous avez les signes.
On va le mettre dans les notes de la show, ou quelque chose comme ça.
Mais il y a beaucoup de lispes et de brises curieuses et tout, en ce moment.
Et les frames principales sont réelles, encore aujourd'hui, je crois.
C'est vrai.
Mais mon préféré de votre avis est que l'industrie aéroport ne compte que 9, parce qu'ils ont
100% de temps.
Je pense que c'est vrai.
Ils ont 100% de temps.
Non, 9.
Donc, il se trouve que c'est possible, mais seulement si vous avez payé beaucoup, beaucoup,
beaucoup de temps et du temps, et faites un set très limité de choses et avez un
grand nombre de complications, des plans et des choses comme ça.
C'est un background cool.
Mais je pense que vous ne faites pas de flight anymore.
Est-ce que c'est vrai, Adam?
Oui, c'est vrai.
J'ai aimé Google pour construire des entreprises.
La première qu'ils ont construite, c'est avant qu'on aille à l'LLMS, et on a juste
fait des mathématiques.
Décadres auparavant.
Décadres, décadres et décadres, 100 ans auparavant.
On a fait des choses dans le mobile, on a construit des machines de processing de temps
qui sont vraiment bonnes et on a vendu ces à la sécurité.
On a fait un très cool « Threat Intelligence Security » pour qui j'ai aimé
une autre company, qui s'appelle Bit.io, qui a fait un service postgres.
Le premier à marketer un service postgres.
Tout le monde aime les choses serverales.
Tout le monde aime les postgres.
C'est super fun, on a fait ça depuis beaucoup de années, on a fait des data bricks.
On a fait des data bricks, on a fait des data bricks.
Maintenant, je suis en train de faire une company, qui s'appelle Market Street,
où nous aidons les petits-business-owners à avoir une plateforme
pour communiquer et collaborer.
C'est vraiment loin de tout ce que j'ai jamais fait.
Ce n'est pas de l'infrastructure.
J'ai fait un joke quand j'ai mis le mien à la fois.
Je ne devrais pas faire des websites, je devrais seulement faire des tools commandes.
Mais je suis là pour faire des websites.
Je ne sais pas ce qui s'est passé, le monde est top c'est terry.
On a fait des websites de nouvelles, ce n'est pas différent.
Oui, ça ne devrait pas.
Tu as juste demandé un LLM pour toi,
parce que tu n'as pas de la façon de le faire.
Donc, tu as utilisé des types de services,
il y a des Next.js, et tout ce genre de choses.
C'est un peu de l'avance, et la technologie de website.
Je l'ai fait juste le jour d'autre,
j'ai vécu mon propre site personnel,
que j'ai été en train de faire depuis longtemps.
Je me suis dit, yo, computer,
je vais chercher l'internet pour mon nom et les vidéos de moi,
et faire un site pour ça.
Et ça a fait tout ça, et ça a pris la frameworks.
C'est génial.
Tu dis ça en internet.
Oui, j'espère que ça marche.
J'ai fait ça aussi, Steve, c'était génial.
J'ai aussi votre site.
La vibe est bien,
j'ai dû apprendre à penser plus grand.
J'ai été décomposé dans mon tête
le problème et je voulais faire un off page.
Et puis c'est comme, j'ai demandé pour faire le tout,
et ça m'a fait le tout.
Et puis j'ai utilisé tous les aspects de mes années
d'expérience au débat,
les problèmes que l'LLM a fait,
et c'est comme, j'ai déployé,
et j'ai commencé à lancer.
C'est comme, ça se passe,
parce que tu es juste en train de faire un type de script
et les gens commencent à dire des trucs comme docker
ou des containers.
Ok, je pense que tu as juste fait le premier,
comme, pas dans le bois ici,
ce qui est cool.
Donc, un peu de automation,
tu as quelque chose de rien,
et puis tu prends ton visage,
et tu as un peu de clôture,
et tu es comme, oh, wait, c'est beaucoup,
quelque chose de mauvais, c'est beaucoup ici,
c'est bien,
mais, et puis un peu de mais.
Donc, dis-nous comment ça va,
où nous sommes,
comparé à ce que, 5 ans plus tard,
c'est miraculé,
et il y a un whole lot de caveats,
mais tu sais des trucs,
donc, dis-nous des trucs.
Oui, c'est bien,
je peux donner un exemple réel,
donc, on est en train de,
le truc que je fais maintenant
intervient beaucoup de data management,
pour faire sure que les LLM
puissent dire des choses intelligentes,
et le déployement de ce genre,
ce que nous avons appelé,
ETL, direct de l'esclosage,
l'exécution,
les trucs,
nous avons déployé sur Vercel,
parce que nous avons utilisé V0
pour nous aider à construire,
nous avons utilisé Curser,
nous avons utilisé Vercel,
et puis, je ne peux pas avoir
longs de fonctionnement en Vercel,
et c'est quelque chose que,
je pense, si tu ne sais pas
tous ces mots que je disais,
et tu ne sais pas ce que tu fais,
je vais changer,
je pensais que c'était
mon provider de base de base,
donc, je suis changé de ceux,
un autre,
un autre,
c'est que tu ne peux pas utiliser
l'Io sur AISinc,
nous nous battons,
quand tu fais ça,
parce que nous avons 95 procs
dans le monde,
et ils ne sont pas transactifs,
donc, nous nous avons réveillé,
et puis,
on a déployé sur Fly.io,
en plus de Vercel,
c'est comme une nightmare,
et toutes ces technologies
sont belles,
elles sont très bien,
mais,
comment, si tu ne sais pas
ce que tout ça est,
et tu as spent longtemps
et une vie,
tu as interprété ça,
même en disant,
les LLM,
c'est comme, passer
l'Io sur AISinc,
postgres,
ils sont comme,
je ne sais pas ce que c'est,
tu commences,
tu te parles à un point
où ça ne te connait pas,
et là,
c'est un peu spécifique.
Oui,
beaucoup de ces plateformes
ont des constraints,
ce qui est bien,
et ça les aide,
tu sais,
à bien exécuter,
mais si tu ne comprends pas
ce que c'est,
tu ne sais pas
ce que tu veux,
ou,
plus important,
c'est de l'interpréter,
quand ça ne marche pas.
Oui,
je pense que c'est
une skill critique,
et c'est un gap intéressant,
je ne sais pas
comment les gens
vont s'y arriver,
mais l'infrastructure
est très difficile,
en plus,
tu as des rails
qui ne s'enchaînent pas,
ou que tu n'as pas
les choses que tu as
à l'expect.
Je ne sais pas,
mais sur le côté de la flèche,
je pense que,
en manière de la perspective,
en beaucoup de façon,
tu ne t'es pas carenée,
si ça marche bien,
et que tu as un produit

tu ne t'es pas inquiétable
pour les problèmes,
et que tu peux s'y aller,
et que tu peux s'y aller,
donc je suis un peu
un peu curieux,
je sais comment se solider,
je vais juste se dévier,
et deux jours,
je vais aller,
et je vais dire,
pourquoi je fais ça,
mais ça doit travailler,
et si tu ne te souviens pas
de la signification,
je pense que ça marche.
Je sais que,
parfois,
les essais sont
en train d'éliminer
le dîner technique,
mais je crois
que,
je vais dire si je suis wrong,
mais un peu de travail
de start-up
est
envers le dîner technique,
c'est juste
poursuivant,
et c'est-à-dire,
un problème
dans les essais,
la tête de la tête,
ou comment ça marche?
C'est ça,
initialement,
c'est beaucoup mieux,
mais ça a pris un temps
de apprendre,
et maintenant,
je m'appelle le dîner technique
d'éliminer le dîner technique,
pour que tu aies un gain,
tu es un dîner,
un dîner,
et ça a un intérêt,
tu veux que la métaphore
soit très forte
dans ta tête,
mais
il n'y a rien de mauvais,
en fait,
dans le start-up,
tu es envers le dîner technique,
donc,
si ça ne marche pas,
tu déclare la bankruptcy,
et tu le déclare,
et tout ce dîner technique
qui t'a fait,
qui t'a fait,
ou d'autres,
c'est que les utilisateurs
sont plus de la code perfecte,
et la meilleure,
en fait,
tu t'es préoccupée de le marché producte
que d'autre,
et de la grossesse,
plus d'autre,
et donc,
le dîner technique,
ça ne marche pas,
à l'heure de la fois,
il y a des aspects de sécurité,
tu es comme,
oui, je ne fais pas ce qu'il y a,
mais je vais changer ça,
mais tu es en train de
faire un carton de dîner technique
du dîner technique,
parce que si tu es successful,
si tu trouves le marché producte,
tu le déclare,
et je pense que c'est vrai,
dans beaucoup de cas,
il y a des exemples,
c'est comme si tu es
construit la softwares,
ou si tu es construit la cryptographie,
et tu es en train de faire des trucs comme ça,
ou des bits d'aile,
où on fait des databases,
on doit avoir des semantics d'acides,
si tu n'as pas des semantics d'acides,
les gens s'en font un,
ils se font un,
c'est un grand truc,
ou un grand important,
il y a des cas spéciales
où tu es effectivement
vendu de la responsabilité
de la sécurité et de la sécurité
de ton marché,
et ça,
tu dois être correct,
mais quand tu n'es pas,
tu dois apprendre
à ignorer tout ça,
en quelque sorte.
Je pense que,
pour le médecin,
si tu es joué,
le dîner technique de la technologie,
ça aide à comprendre
que 1 000$ de dîner
versus 1 million$ de dîner,
quel est le currency que tu es en train de utiliser?
C'est 1 million,
je ne suis pas sûr.
Oui, c'est un grand point.

Est-ce que ce dîner que nous sommes en train de créer,
est-ce que c'est quelque chose que nous serons
d'être,
en pensant que nous nous succèdent,
est-ce que c'est quelque chose
qui va nous ruiner
en post-acquisition,
ou juste avant l'acquisition,
ou peut-être un peu plus,
quelque chose comme ça?
Donc,
peut-être que tu dois avoir
tes yeux à l'opinion de cet dîner,
en termes d'être capable d'understand,
et de ne pas juste en tourner,
mais tu le réapprends,
tu le marques sur le table
et tu le taux.
100%
Je pense que part de ça,
c'est la management de risques
et de la manière dont ça s'applique.
Le autre,
c'est que tu as les mêmes metaphors
qui sont sur le côté de la flèche
de la même métaphore,
c'est que tu n'as que
que tant de tokens d'innovation
pour les expérimentations,
et tu veux vraiment
les expérimenter
sur les choses
qui font le marché du product
et qui servent à vos utilisateurs
ou à vos grosses,
ou tout ça,
comme tu fais,
et tu n'as pas envie de les expérimenter
en,
en,
en movant tes outils
et vos cloud providers
17 fois pour un week-end.
Et pourtant,
c'est ce que j'ai fait ce week-end.
C'est comme,
comme,
quoi est la expression de
comme,
des coveillards,
les enfants ne ont pas des chaussures
ou quelque chose,
mais tu les en as,
tu les en as,
je ne les laisse pas souvent.
C'est un problème.
Donc, un second anniversaire,
tu disait que tu n'as pas de marché,
je ne sais pas,
j'ai écrit un article
sur lesquels les gens
utilisent la AI4
et des idéations productives.
Je sais pas,

le process de souci,
un processus d'infermentation très tard dans l'infancy de la start-up, ou même avant
que vous puissiez avoir ce point.
Alors, qu'est-ce qu'on veut faire ? Ou qu'est-ce que vous avez un kernel ? Et avant
que je fasse un peu d'ingénieurs dans un sondage, ou même que je fasse un peu de code
ensemble, juste avoir le concept de l'AI, un peu de websites, faire des idées, et puis
vous avez les gens avec des poignées de la poignée qui regardent ces choses.
Non, je ne peux pas comprendre pourquoi vous voulez faire ça.
Vous avez vu maintenant une iteration beaucoup plus rapide sur les idées de la poignée
avant qu'ils puissent avoir des ingénieurs à leur temps ? Ou est-ce que nous
avons maintenant vu les choses construites ? Et que nous sommes plus fortes ?
Est-ce que ça va se faire mieux ou plus mal ?
Oui, est-ce que nous sommes plus fortes à des idées malades ou nous sommes plus fortes
à construire ? Ou est-ce que nous sommes plus fortes à la vécu ?
Oui, oui, oui, oui, oui.
C'est ça.
Oui, oui, oui.
Alors, je pense que mon tournage sur ce point est un peu différent de certains des gens.
Il y a des founders et des CEO et des entrepreneurs qui sont incroyablement passionnés
par un produit spécifique.
Ou un problème, je dois dire.
Les gens sont...
Ils sont...
Mon vie est créée pour faire de la santé, ou quelque chose, et j'aime ces gens.
Ils vont faire leur chose, en ce moment.
Je suis beaucoup plus de...
On analyse le marché, on arrive avec un peu de idées durant la phase de la phase de la brainstormation.
Moi et la personne que j'ai construite, les deux premières entreprises,
nous avons un spreadsheet de 150 idées.
Tout de suite, et nous avons juste construit pour 10 ans,
et tout est sur la table.
On a des idées très belles, ce que je appelle des idées belles,
c'est que, si on avait des idées de la maison, c'est une bonne idée.

Et puis, évidemment, pas...
Ouais, on le bat, on le bat.
Ou des idées pratiques, comme, vraiment,
les post-crash sont sur la table, et c'est ce qu'on a construit.
Et puis, nous continuons de mettre des colonnes à ces roses en chute
qui sont toutes les critères de la exclusion ou de l'inclusion.
Et pourquoi je vous le dis maintenant,
c'est que ce que l'AI a vraiment changé pour nous dans la phase d'idée,
ou cette idée, mais c'est le cool, le petit de la façon de dire ça,
à ces jours, est comme,
qu'est-ce que l'AI a fait pour les vêtements?
Donc, pour nous,
avec l'AI, vous pouvez construire beaucoup plus vite
avec tant de plus de ingénieurs,
et vous pouvez avoir une phase de prototype,
et vous pouvez tester si ça fonctionne tellement vite.
C'est bien, parce que vous avez le temps de trouver
si vous avez un marque de producteur,
vous avez le temps de faire toutes ces choses
plus cher et moins d'argent,
mais ça veut dire que tout le monde
a cette habilité aussi.
Donc, votre mote,
la chose qui protège votre compagnie,
votre idée, vous avez besoin de trouver de nouveaux motes,
que vous avez précédemment,
c'est comme, nous sommes juste un peu de technologiques,
nous avons construit de la technologie,
et nous avons écrit beaucoup de codes,
et c'est notre mode,
c'est notre propre intellectuel, etc.
Mais maintenant, quelqu'un d'autre
peut être comme,
yo, vibe code, me insta-card,
et ils font,
et le mode est beaucoup différent,
ces jours, et donc vous devez trouver de nouveaux aspects.
Avec ce que je suis construit maintenant,
les aspects de la communauté et de la networking,
entre les gens,
je pense que ces connecteurs sont persistants,
et que l'IQ a des aspects,
je pense que vous regardez le long terme,
les things long terme,
les things valuable,
les things have that side of community,
or they're like 10x technologically better.
And the LLM's not going to build something
10x technologically better,
and it can't do the community side of it,
but it can do so many other consumer apps,
or other, or SaaS, infrastructure SaaS,
is a great thing that like,
the LLM's can do that, right,
and they will build that,
and it'll be fast,
and so I think it changes the dynamic,
I think you do get a lot of people spamming,
I added AI to X to the VCs,
but I think the smarter VCs are saying,
like, how is this going to possibly
100x my investment
when anyone else in the world
can just ask the LLM to build the same thing.
And so you need to have a different lens,
I think, when you're choosing your ideas,
and it has to be that lens of like,
well, one, it has to have a lens of like,
is this just a Giro ticket at OpenAI,
or in Gemini, right,
like, is this just going to happen in two years, right,
like is this some like product manager
who's going to do it,
and then also like,
what makes this sort of AI proof
or what builds a moat
that defends it against other people,
you know, vibe coding it as well.
So taking us to the general theme
of our podcast here.
Yeah.
How does it mature?
How does the unlock the achievements of
requires reliability engineering now?
Like we go from thing with slide deck
to thing with random website
with some, you know,
we bought some time in a cloud to,
okay, we have some employment,
we have people building stuff,
and now there's stuff that people use
and they expect them to stay running.
When do you decide
it's time that the,
we don't just have stuff that's running,
we have actually people accountable for running,
and we have expectations.
What have you seen that?
We have to growth now.
Yeah.
Tell us about your like long arc
of experiencing this here
and in your own industries.
Yeah, I think and Bit.io
was a good example of this where like
Bit.io is serverless Postgres,
so it has to work, right?
But it also has this unique thing
where with serverless Postgres,
you, the goal is in a way
is to shut the machine now,
shut the compute down
when people aren't using it, right?
But it's a database.
So how do you do that?
How do you do that safely?
How do you do that without corrupting data?
How do you do that like
with transactions in flight?
Like what do you do?
Time them out?
You know, you have to start thinking
about those things,
but when do you start thinking about them?
You start thinking about them
when you have what I kind of consider
as like a successful failure,
if you will, right?
Like if some user using you
and they're like, yo, this thing broke
and they're mad about it
because it means something to them.
But hopefully it's not like,
it's not something you can
necessarily anticipate or see
because you should deal with those
if you can't see that.
But you want to have enough load.
You effectively want
the Twitter fail well, right?
Like Twitter did an amazing thing
of turning the fail well
into like marketing, like genius.
Like I don't know how they got that done,
but genius.
And then you say,
we have to spend time on this
because it matters
or we're losing customers
or people are noticing or whatever.
And I think that,
that isn't to say,
I mean, I think if you can,
you should plan it well.
You should take care
of building your software to plan it well.
You should know where your problems
are going to be like the
back to the credit card thing
of technical debt.
But I always think about it as like
when my users notice
or complain
or if it's some fundamental problem
in the entire
like product I'm offering,
like is reliability,
like we said earlier,
is reliability part of my product?
If it is, then yeah,
we have to be,
it's got to be built in.
But man, you got to wait
until the user tells you
something's wrong.
Because if you don't,
you'll build the wrong thing.
You'll invest all this time
in fixing a problem that doesn't exist.
So if we can
follow the model
that you're describing here,
like let's say we have a,
you know, a couple of startup founders
who don't have a reliability background,
don't have a security background
and they're not building a reliability
or security product.
They're building a consumer product
or whatever.
They're building Instagram too
or whatever.
At what point,
like what's the switch
where you think a founder has to be
like we need an infrastructure person
or maybe not infrastructure
but a reliability person?
Like at what point
do you make that jump?
Reliability or security person.

And I think that
when you start seeing the growth.
Specialist.
Yeah.
I think when you start
seeing the growth,
when you say,
when you can sort of project
your growth curve out
of users or whatever,
you know, in the consumer app space,
what you're saying,
like, you know,
like if this keeps going this way,
right, like we are going to run into these
problems, like you should make those
projections and say,
and it can be hard to know.
Or, like you said,
in this case,
if you have like a founder
who's not a SRE
or a security person
or just doesn't have exposure
to like what can fail
and when it goes wrong,
I think that you should still
sort of say to yourself,
we're growing so fast.
How does this impact
what I've built?
Like how does this,
like can the choices I made,
which in a modern startup,
right, you're composing it
of SaaS services,
like a modern startup is like,
like I said earlier,
like you use Vercel
and Fly
and Superbase
and Neon, whatever, right?
Like you compose all these other SaaS services,
which is not traditionally
how you do it, right?
Like usually it all
be in one cloud provider
and you may need to go
in that direction,
you may need to go,
I think eventually one of the reasons
the major cloud providers
are so successful
is that they offer such a coherent
and built in policy around,
they offer you this path to scale
and this path to security.
With the downside of the complexity
that comes,
but I think it's,
you just always have to be looking
at your growth and say,
question your underlying
building blocks, do they work?
Are they big enough?
Are they strong enough?
It's hard.
Another thing that you mentioned was,
and maybe I'm projecting here
a little bit, but like it sounds
like the story you're kind of getting at
is also like the failures
that you're going to encounter
when it comes to scaling up,
like you can't actually see them coming.
Like some of them,
maybe you're like,
well, last time I did this,
it was when we got to a million
concurrent sessions per whatever,
but like now you're on fly
or whatever, right?
You're on, you know,
you know, service.foo
that has never existed before.
And so maybe their ceiling
is much higher or much lower.
And so there is this community
out there called Learning
from Incidents.
I don't know if you've heard about this one,
but it's basically like,
I know when I first heard about it,
it totally made sense to me,
even though I hadn't heard
their actual background,
you know, hero story or anything,
but it was like the way that you
scale things over time
is that you fall occasionally
and you learn when you fall
and you're like,
aha, that's where the line,
that's where the bump is.
We found the bump, right?
And the trick is to be able
to fall kind of gracefully
to extend the metaphor a little too much
to only fall, you know,
in parts of the globe at once,
potentially, if that's possible,
you know, like in some of your user base,
you know, so we get into like the
continuous deployments,
but across failure domains
and like a gradual rollout,
kind of a blah, blah, blah.
So like, I guess what I'm getting at
is like in that modern infrastructure
world that we're talking about,
like how many of these principles
transfer effectively?
Like you and I worked on board together
with GSLB and all the things
inside of Google,
like do these words,
I mean, not those words exactly,
but like do these concepts
translate into this like today's
operational world
and do people get it?
Or like, what's the,
is there a Rosetta Stone
that needs to be built for this world?
Yeah, I think there is,
there's definitely a Rosetta Stone
that needs to be built from Google out.
Yeah, if you're a Google
and you want it,
yeah, and not encouraging
any Google Clips to quit,
if you want to build the startup,
I think you do,
you will need a Rosetta Stone, right?
Like the way we built software
Google is very different.
However, I think you have things
like the SRE book,
you have that just the simple thing
of like, ask the LLM,
ask Gemini,
like, hey, I have this,
I am deploying on
Vercel with Next.js,
my user rate is growing,
like where am I going to run into issues?
You can just literally ask that question
and it will give you good answers,
right, because a lot of these
best practices are well codified
that the LLMs know,
and they'll be good enough
to get you to the next step.
And then you can literally,
you know, it'll start saying things like,
you know, often scaling comes like,
check your database, right?
Like, are you on some free tier
of some sketchy database?
Like, maybe don't do that.
Like, do you have indexes, right?
Like, and the big gap to me would be,
we still, like, I talked to a lot of
founders, technical founders even,
and they just like,
like that idea of like create indexes
in your database.
That idea has been around for 50 years,
right, like, I mean, literally forever,
right?
And that is still like mind blowing
science to a lot of people, right?
They just don't,
they don't think about those problems,
right?
They just assume that it does it all
right for you.
So, I think the luxury we had at
Google or have it, you guys have at
Google, is that a lot of those things,
you know, one, there's very smart
people, like you guys, to answer
those questions when they come up,
but also there's great frameworks
to handle it, and we've thought a
lot about it.
But we thought about it in the
context of the infrastructure Google
has internally built, which is
very different from what the rest
of the world is like.
You don't get those things.
You don't have GSLB.
You don't have, you know, you don't
have, even down to like the disk
infrastructure of like D, and like
all the way up, like to big table.
You don't have those things,
colossus, like you don't have any
of that, and you have different
assumptions on how you things are
built.
This really brings to mind
something, like two things come
into my head.
One is, first of all, you've done
this sort of meadow heroes journey
a few times, which is kind of
fascinating to me, which is going
around the clock with multiple
times of, okay, first you have
nothing, now you have toy website,
now you have toy website with
super power X, and then let's put
thing on cloud, oh, never mind,
scrap that, let's do it again.
Like, okay, now you have a magic
flute, and now you're going to go
in a castle.
Okay, so my question is, the
actual question is, so you take
some off the shelf, perhaps, you
know, open source tools, then
migrate to possibly paid sources
made by third party, then you go
mmm, actually that thing can't do
a thing I need, because it won't
scale.
What does that like to migrate
from paid tool, sort of does
what you want, now you're going
to do it yourself.
Oh, you're going to, oh yeah,
no, that's a pain in the ass.
What's that, what's that, you're
doing it yourself part like,
because it's going to, it's big,
because you're going to do it,
it's got to scale up for the
thing you're doing, and now
you're writing yourself and you
have to support it, and my
guess that's where the quote
fun happens, and scaling and
the reliability part becomes
scary and opportunity.
That I find really fascinating.
Yeah, and I think you're right,
I think there's like these
influx, but now, for me,
personally, and for the last
co-founder, my current co-founder,
like, we come from this
background, so I'm like excited,
like, I'm like great, I get to
do cool shit, and then everyone
who's ever worked for me
reminds me, it's like, no,
you don't, you have to go talk
to an investor, right?
Like, you're not allowed to
build any of those things.
But the migrations are scary,
first, right?
Especially if things like
database migrations are
fundamentally, like, the
stateless web migrations can be
fairly straightforward, but
migrations between database
providers, authentication
providers, things like that,
migrations from a, like,
alternative cloud to, like,
an honest cloud platform, like,
AWS or GCP or Azure or
whatever, right?
Like, those are hard, and
they're hard for a lot of
interesting and, and that's



fun reasons.
Like, if I never have to deal
with IAM and policies again in
my life, like, I would be,
it would be like happy, right?
Like, I'd be in my happy place.
But you have to deal with them,
and they're there for a reason
or, you know, like, try
explaining, you know, like, oh,
well, you need a private link
between these two VCPs.
By the way, you want to be
multi-cloud, so like, those
terms don't apply in other
clouds, right?
Like, how do you set up a
redundant VPN link between
two cloud, like, yeah, you
hire a specialist, like, this
is an area where you,
that's very, very hard.
And so the reason these
alternative clouds exist for
people like me or people who
are building startups is
because they want it to be
easy.
Bit.io, the whole premise of
Bit.io was like, it sucks
to run databases, don't ever
think about it again.
And we will tune it, we will
do the right thing, we will
scale it up and scale it down
for you, like, we'll do all
of that for you, like, and
that was successful, right?
People love that, they still
love that, there's neon and
super base exist because of
that.
But yet, eventually, yeah,
you, if you gotta move
sometimes.
Sometimes you have to move
because someone buys your
company, and all our people
at Bit.io, I apologize
when we got bought and we had
moved databases, that sucks.
Sorry, I hope we helped you.
Yeah, those transition
points are hard.
They're very, very hard.
And they're still hard.
And no one has made that easy.
Doing the classic podcast
thing, we now require you
to predict into the future.
Great.
So good at that.
So projecting from what we've
been talking about for the
next one to two years,
like, what else do you see
that will become weirder
or less of a pain in the
butt or something like,
like, how does this arrow
point forward?
It's going to be significantly
easier to create software.
And that, actually, I love.
Like, I've spent a career
trying to build tools that
make it easy for people
to build software, like
developer tools and things
like that.
Like, one of my big things
is like software is the most
amazing thing that we've come
up with, I think of a problem
and I solve a problem.
Amazing.
And I think it's just going
to be easier and easier
to do that in two years.
You're going to have
hyper specialized, hyper
quick builds.
You're going to have a lot
more internal software built
to solve internal problems
that will be very customized.
I think you're going to have
a lot less need for software
engineering in the
traditional sense.
You'll have more product
people build software
using the vibe coding stuff
and all that.
And that will all be net
good initially, I think.
What about some sort of
mutual, some happy coexistence
between us and the robots?
Yeah, I definitely think
that's possible.
I mean, the whole premise
of the company I'm currently
building, Market Street,
is that the human connection
is the important part, right?
Like, small business owner
to small business owner,
those people will still exist.
They will have, you know,
they're doing real problems
talking to real people,
like they're working on
those spaces.
And I think that AI
can complement a lot of what
they do and help them
grow their business
without being
extractive against them
and can help bring to them,
you know, help build
human connection, right?
And help build human outreach.
Like, that's my goal
and what I'm doing now is,
and the driver of that
is sort of focus on the humans,
not the computers,
and focus on the human connection
and focus on the interplay
in neighborhoods
and between people.
And I think that's actually
like where we need to be
spending, obviously,
and betting on it,
but I think that's actually
where we need to be spending time.
There, of course, needs to be,
you know, building the technology
that gets us there,
but I think that people don't go away,
right?
Like we have to work with each other
and I think the human connection
is the non-replaceable thing
for AI.
And we need to focus on that.
Yeah, and on an optimistic take,
like one of the kind of lines
that we use in sort of
the resilience community
is like a, it's a socio-technical machine.
So it's not just,
not just the machine driving,
but there's people in the loop, right?
So we can always,
I don't want to say,
hope on, count on.
You know, hope is not a strategy.
Hope is not a strategy.
It's not a thing.
It's not a strategy.
But like, you know,
there are people to,
at least assist in making decisions
and steer things and observe
problems and move in the right direction.
I want us to go there.
I want us to be more creative
via artificial intelligence.
I really want this to be the time.
Like we always say
when technology comes
and we bring automation
and that like that frees people up
to have their time
and their needs taking care of
and like create it.
I really want that to be true.
So far, it's, sometimes it works.
But a little more tangible.
I want to bring us back
to another point.
So specifically,
what I,
if we skip the big philosophical
angle and strictly focus on
sort of the sasses of the world
and the passes
and the whatever's of the world,
do you have like,
the reason why I'm picking on you
is because like,
you're necessary who is like
further towards the edge
of this technology
than most people that we talk to.
So I'm curious what you think
are the problems with the current
offerings that you think might be
solutionified
and turned into platforming things
next.
Or what's your wish list really
in terms of how to progress this stuff
so that like the next round,
you know, Adam, you know,
3.0 will come around behind you
and like have a better time
solving problems like this.
What I really want
and people have heard this
ramp before,
but what I really want is
the cell for as a service,
the platform as a service
that is the SRE model from Google
to exist in the world, right?
Where, you know, at Google,
Oh, shucks.
Oh, yeah, I know.
Well, we have ownership, right?
Like, I think that every time
we had a product that we
onboarded into SRE,
it was a paired partnership
ownership with product
and with everybody involved, right?
Where SRE had the full
and expectation of modifying
the code, building the code,
making it every,
like solving the aspect of reliability,
often to security
and things like that, right?
Like, would come in
and we were just like,
hey, we've changed your code, right?
Like, literally, I think
one of the first projects
we're going at Google
when I stopped doing airline stuff
was Google News.
And we rewrote, we moved,
we ported Google News
from the old framework
that was onto a framework,
internal framework,
Google has called Apps Framework,
which may be still exist,
I don't know.
But it offered, like, underlying it.
It was like an SRE-led initiative
that underlying it was all these things
like load shedding
and, you know,
all these different reliability things
that made it just much,
much easier to operate at scale
and much more reliable.
And we did the port
from the sort of bespoke
server like...
We, the SRE team.
We, the SRE team did the port
over to Apps Framework
and it made Google News
a much more stable
and reliable platform
and helped to grow,
you know,
or helped stay alive
and grow from then on.
And I wish we would see that
in the world today.
Like, we get all these people
building SRE,
quote, SRE SaaS products
that are not that.
They're, like,
very reactive.
They're very about monitoring
and they're very about,
like, all these other things.
And they're all reactive.
They're all like,
oh, we reduced your MTTR.
I'm like, well, why don't you
just change the code?
Like, I want to check in my code
and I want it to be like,
yo, I added all this stuff.
And by the way,
all of it's awesome now, right?
Like, all the reliability stuff
is awesome now.
Like, you weren't doing multi-region.
I got it configured.
If you want it, here's you turn it.
You weren't doing load shedding.
I got it.
I turned it on.
Right?
Like, that's what I want
from my next SRE SaaS.
I don't want,
I don't need more monitoring
and I don't need alerting
and I don't need, like,
incident management
or SLOs or, like,
all that.
I have all that.
Like, that's par for the course.
Isn't that requisite,
like, at the code layer?
Like, it's not at the platform
infrastructure layer.
Like, it's in the code.
100%
Yeah.
It's in the code.
This isn't a platform,
you know,
you know, hot swap Indiana Jones situation.
This is a, like,
you now need to build the thing this way, right?
You're going to use this framework.
Yeah, there's a migration event,
but, like, there's a lot of,
I mean, I work with customers a lot,
like, migrations don't happen,
honestly, very often.
Like, they are very...
You don't want them to.
They're costly.
Yeah, yeah, they feel
to be zero value often
or negative value, potentially, right?
Yeah, you're now introducing
a maintenance burden
to your development team.
That's going to be,
well, why am I doing this?
And, like, well,
your uptime is going to go up,
and, like, I,
and I now don't understand
this new platform I'm on yet.
Right, right.
I look forward to understanding it
in the future,
but right now,
it's just the cost of understanding
this is very high
and your reliability team is,
like, look, it's going to be great.
It's, like, but I haven't been convinced
and I have a bunch of new features
I want out next quarter.
And that sounds to me
to be a discussion that we keep having.
And, like, if you are a big,
big company,
you can afford that cost.
And your developer team
will give you a lot of, you know,
a lot of rope to work with.
But if you're really small,
that is a really difficult decision
to make for your team,
especially if you're working
with a limited budget.
And I agree with you
if these things were in place
all around you,
you don't need to figure this out yourself.
We've already built you
a resilient framework.
Just kind of use this one.
It has load shedding in it
and monitoring tools in it.
And, you know, alerting in it.
Why would you do the other thing yourself?
That would be madness.
So...
Yeah, and I want the tools,
the platform to effectively do
the migration as your downtime
for me, right?
Like, I want it to be, like,
here's the Git commit.
By the way, when we can run it,
it will split, you know,
you'll have an A-B test,
we'll do it, like,
and it just does,
like, I don't have to think about it.
Like, a very practical one, right?
Like Django, you know,
it's an ORM for Python
and a website-building tool for Python.
It's very, very reliable.
It's been around forever.
There's, like, migrations in it.
But there's also, like,
you have to run it, like,
inside another framework.
They're from, like, Unicorn
or UVcorn is the new one, right?
And that's the Async version.
And, like, what are the benefits of that?
How do we explain that?
How do we automatically port you over to it
if you're in a legacy Django?
It's, like, you should just do that.
You should know.
It should be, like,
and have installed these extensions
and they're configured and they're running.
And we tested it here, right?
Like, and that's what I think, like,
if someone came to, you know,
Google SRE, they would own and do that.
And, of course, it's a little easier
because the SRE org is convinced
of the value of that, right?
Right, right.
And I think performance,
for what I think performance is the magic
answer to a lot of these questions.
And it was at Google, too,
like, when we would pitch things,
people understood the value of a millisecond, right?
Like, you know, like, I used to go visit the GFE team,
the Google front-end team in Cambridge
when I was there and I go downstairs
and, you know, they're ready to go.
And they had this big diagram on the whiteboard
that was how many nanoseconds
each branch of the GFE took, right?
And, like, because they cared, right?
Like, that mattered a significant amount, right?
You're just a piece of infrastructure
that can't take, your budget is not milliseconds,
it's nanoseconds.
So, like, but performance results in delight,
it results in the user, right?
And it results in, I think,
you can actually, like,
track the value of performance, right?
And so, one thing you can say is,
oh, we're doing this migration
not just for reliability,
but also look at this cool thing,
this performance thing,
which actually increases your bottom line.
It's a revenue generator,
it's not a cost center,
and we give you clarity as to why it matters.
I think that's, like, one of the hacks you have.
Yeah, there's a similar story
from years before around news where,
but this was in, if you remember,
there was, like, a downloader service
for, like, apps that Google provided
for, like, Windows and Mac and stuff,
like, I think.
Oh, yeah, yeah, yeah, yeah.
It was just, like, the blob serving,
but it was, like, one specifically
for, like, applications,
so they had, like, versions
and package manifests and blah, blah, blah, blah.
And it was, like, homegrown
for, like, years and years and years,
and one very, very cool software engineer
who doesn't describe himself
as an SRE-win in,
and was, like,
I have replaced you with a small Go script
because I, like, go a lot.
And, like, the reason why it worked
was performance went, you know,
through the roof, simplification of the code,
like, it dropped dramatically
and, like, they were able to shed all of the features
that they didn't really use anymore
and, like, no one would have done that on their own,
but, like, he was just, like, a senior engineer
who was, like, looking for something cool to do.
And, like, that was, like, I feel, like, that value,
if we could package that concept
and bring it out to the world and be, like,
this is actually good.
Like, when executed well.
Like, you don't want someone spending 16 years
trying to do that with a team of 55 people
who constantly cycle and waste a bunch of money,
but, like, if you could atomically do that thing,
you can get a huge amount of value out of it,
not just in terms of money and performance,
but, like, reliability and probably security
and all sorts of other things, like,
maintainability and all these things.
So, that's really hard to get across to everybody, I think.
I think you're right. I think it is very, very hard to get across.
I mean, I harp on performance as this.
You know, one of the reasons is because you can always
show that cool graph, man, or, like, it's, like,
this took 200 milliseconds, now it takes 10.
Right, like, those graphs are very convincing.
And this is true for beyond the startup world.
This is, like, true in the, you know,
the small business mediums,
the big business world, like, man, those graphs sell.
Right, like, you've really changed a lot
and you can point to the direct sort of monetary impact
of those less compute, less memory, you know,
or maybe not less memory, maybe you're doing the trade-off.
But, like, still, right, like, you get to point to it
and I think that that's, like, that's a big, big thing.
And if you include in that reliability,
which often you do, and I think you made a good point
about, like, code simplification and, like,
architectural simplification, right, like,
you got rid of all the things, you got rid of all the assumptions,
you understood the abstractions better, right,
like, a good refactor does that.
And I think that's one of the things I'd love to see,
this Astool start doing that.
I think we're close.
I think we're, like, a year out from, really,
the LLMs being able to do quite a bit of that
with some, at first, human hand-holding.
Yeah, cause that is new.
The ability for an LLM to go in and, like,
actually do, you know, open heart surgery on your code
is new.
Like, it didn't exist, this wasn't here five years ago.
Yeah, it didn't exist, like, no way.
Yeah, it wasn't here like two years ago.
I mean, I, like, the first time I tried doing something
two years ago with, like, co-pilot, I was like,
wow, you can't even do math, right?
And now I'm, like, build my whole site.
It's like, I got you.
Start over.
Yeah, do it again.
Yeah, no, exactly.
It's always, like, oh, you're right.

The other day I was vibe coding and it did something weird.
I literally wrote into the chat window, not like that.
And it did the right thing after that.
It's just the right thing.
I changed it to, yeah, I always feel like it's, like,
reading the first or second design doc
a junior engineer has built, has written.
And you come in and you're, like,
I see where you're coming from, great assumption,
but, like, trust me on this one,
you really want to do it this way,
which is the actual best practice.
And so, like, then they're, like, oh, cool.
And they go do some research.
Oh, yeah, yeah, you do it like an expert.
Oh, okay.
Oh, yeah, okay.
Oh, yeah, you're right.
You're right.
No, no, do it like a world class expert, like, oh, okay.
Oh, good idea.
I should do it like a world class.
It's, like, so sorry that it didn't before.
I'm like, just do it before.
I was trained on the internet, not the expert in it.
Sorry.
All right, well, we're getting close to the end of our time here.
So, I'd like to ask you a question for a number.
A number in your travels with a unit, preferably a surprising number,
like a really big number or a really tiny number
or a fast number or a long number,
something that really is exceptional.
A number with letters in it and numbers.
Yeah, yeah, and something, yeah, something is, like,
wow, that is impressive to you,
that surprised you in your studies in the world.
And you're, like, wow, I found out this thing
when I was looking at a thing, it was, like, I had no idea.
This is a really big thing,
or this is what people need or want or whatever it is
that you would like our listeners to hear about.
That's something that really impressed you.
Oh, boy.
Okay, I have a sort of funny anecdotal story about numbers
and that everyone relates to my work at ITA.
Is that if you know the confirmation code you get on your ticket.
Yeah, letters and numbers things, yeah.
Letters and numbers, right?
And the reason those are letters and numbers is because, yeah,
is that this is not a confirmation code.
This is, like, the flight, our first flight we took
on our reservation system.
Close to it.
Anyway, confirmation code, those, like, six characters.
The reason those are that is because when they first started
storing your tickets and record locators on disc,
that was the block address on disc, physically.
On a cylinder somewhere.
It's a physical pointer.
So it's, like, literally, yeah, like, literally, like,
a physical pointer.
So they just wrote, it's like, imagine you print F to the pointer
and you got the address and you just, like, here, this is yours.
Is it?
This is how you find it.
Are those locator codes?
They're not in hex.
I've seen Zs and stuff in there.
Yeah, Zs and, yeah, they're not in hex.
Like, literally, like, IBM 360 disc addresses.
They're, like, some sort of 36 character.
Yeah, some sort of 36 character by six.
Should be fine.
Yeah, factorial six or whatever it is.
I didn't bad a math.
And, like, that number is, that's always, like,
really impressed me that, like, the leakage,
I just talk about, like, leaking your abstractions.
But, like, it worked great, right?
You could find it fast.
Hash tables are the most profitable data structure.
Oh, right.
The railroad grade standard of flight locators.
This is fantastic.
Oh, oh, oh, oh.
Just don't run across a very large key space.
Like, let's do it.
Yeah, let's do it.
Yeah, let's do it.
But now you have to support it, right?
Like, solve the new stuff that doesn't do that.
It's like, what the f***?
Why are we supporting this?
That's awesome.
Yeah, yeah, it's pretty funny.
The other fact that I find interesting about small business,
related to small business and things like that is,
one is every time I've talked to a small business owner
and asked, do you use AI?
They have laughed at me.
Of course I do.
Which way do they laugh at you?
Like, of course I do or of course I don't?
No, of course I do not.
So the people that you, when you walk into a retail store,
like, think about the technological differences,
the age gap, the exposure, the cultural context of all those people
is so different than what we are.
And let's remember what medians mean,
like above and below that number.
So yeah, that's pretty awesome.
Yeah, exactly.
Yeah, a lot of it are a lot of, yeah, obviously 50% above.
But, like, yeah.
Cool.
Well, thank you very much, Adam.
This has been a great talk, unsurprisingly.
Is there anywhere on the internet
that you would like people to hear your words in other ways
or otherwise, like, hear you make jokes
or talk about what runs you go on
or whatever sort of socials you like to share?
Yeah, I'm on Twitter at writeaheadlog.
Very nice.
Which, if you're a database person,
you will find that somewhat funny.
Or perhaps not, if you've ever used or had to repair one.
Then, my favorite PTSD.
My company is MarketStreetMarketST.com right now.
There's a funny story about the actual domain.
I'll tell you later.
And then, yeah, check us out.
Look, email me.
And if you're, like, some sort of SRE
that's starting a small business,
I want to hear from you.
It's, like, strange.
But, yeah.
And computers are the best.
Cool.
Thanks, Adam.
Thanks, guys.
Thank you.
And don't forget everyone that hope is not a strategy.
And we'll see you next time.
Bye-bye.
You've been listening to Podcast,
Google's podcast on site reliability engineering.
Visit us on the web at sre.google,
où vous pouvez trouver des papiers, des workshops, des vidéos,
et plus de SRE.
La podcast est hostée par Steve McGee,
avec des contributions de Jordan Greenberg,
Florian Rathgeber et Matt Siegler.
La podcast est produite par Paul Gulli-Amino,
Sunny Schau et Salim Virgi.
La podcaste est télébotte par Javi Beltran.
Special thanks à M.P. English et Jen Pettoff.
Vous avez perdu le téléphone.