text stringlengths 0 2.29k |
|---|
**Gerhard Lazu:** For sure. For sure. Okay, so I'd like us to come back now to your talk, because one thing which I really liked in that video, and in your presentation, is you talked about why you haven't chosen Microsoft, and why you haven't chosen AWS. It's not like, you know, "We haven't even looked there." You did... |
**Florian Forster:** Yeah, the first and most prominent thing that struck us was kind of having end to end HTTP support. Because we provide gRPC APIs to our customers, and they need HTTP/2. And while verifying that with all the different offerings, it was kind of hard to either get proper documentation, whether they su... |
\[29:51\] I mean, we could have chosen the route and say "Okay, we do not offer gRPC, but only gRPC web and REST, because we supply that as well. But we really wanted to have the HTTP capabilities, because we think at one point there is a unique opportunity to be taken to use streams and stuff for identity-related thin... |
**Gerhard Lazu:** Yeah. It adds up, right? Seven milliseconds here, three milliseconds there... Before you know it, it's a minute. |
**Florian Forster:** No, really, if you have microservice architectures -- I mean, if you have like five services cascaded, and every time they call like our introspect endpoint to validate some tokens, it adds up. It's 50 milliseconds only serialization at that point. But that was just the decision they made there, an... |
**Gerhard Lazu:** Tell us about them. |
**Florian Forster:** I mean, I still to the day not fully understand why Google Cloud Run needs to use their internal artifact registry... Like, you need to push your images into Google's registry, and from there on out, you can fetch it in Cloud Run. I don't know why that decision was being made. There might be a tech... |
**Gerhard Lazu:** That's right. One thing that maybe we haven't done as a good job to convey this is that your service is global. So when you're saying VPC connect, you don't mean just in one region or in one zone; you mean across the whole world. So how many pops do you have worldwide, where Citadel runs? |
**Florian Forster:** We have like a core pop region. It's like three regions we run constantly, that comes down because we run our storage there as well. And sometimes if we see different traffic profiles, we start regions without storage to them, just to get some of the business logic closer to customers. So that can ... |
\[34:13\] We did some internal experiments, we built like a small TerraForm function where you basically can throw in a list of regions you want to deploy, and it will basically deploy to 26 regions in like one to two minutes... |
**Gerhard Lazu:** Wow. |
**Florian Forster:** That works really well. But you get strange problems if you do that, because sometimes you want to have like hot regions, because your application is anyway running, it can serve traffic quite easily... If you have to cold-start a region, it always takes a few milliseconds to do that. And it's not ... |
**Gerhard Lazu:** Yeah, that's crazy. Like, you say that two seconds is slow for a whole region to come up. It's like, "What?!" \[laughter\] Like, try booting something; it will take more than two seconds. Anything, really. Wow, okay... |
**Florian Forster:** I feel it's like engineering ethos that you might at some time over-engineer certain things... But still, it feels right to do that, because it's more easy to just scale up an existing region and throw some more traffic into that. You can easily steer that around. The thing you most of the times wi... |
**Gerhard Lazu:** Yeah. But it's interesting... Sometimes routing to a region which is further away, sending a request to that region can be faster based on your workload. And even though your workloads are super-optimized, for something to be up and ready in two seconds, that's just crazy. Try doing that with Java... ... |
**Florian Forster:** I mean, it's even small things involved into getting like fast startup latencies... One big drive of-- we hypothesize - we don't really have evidence, but we hypothesize - is that the image sizes of your containers influence that quite a lot. Even though I think Google does quite a lot of magic in ... |
**Gerhard Lazu:** \[38:14\] Okay, okay. Yeah, you're right, you're right; that can make a big difference as well. So a few seconds is not bad, right? Especially if we have like a blue/green style of deployment, where -- and I know that Cloud Run supports that. So you're not taking the live version down, and that's okay... |
**Florian Forster:** Yup. |
**Gerhard Lazu:** Okay. Okay. How is that like? How are you finding the Google CDN? Because I haven't used it in anger; I mean, only small projects... How does that work in practice? |
**Florian Forster:** I actually quite like it. One of the things we like the most is that you can cache assets across multiple domains. So for example, each of our customer has their own domain name, and our management GUI is built with Angular, so we have a lot of static assets to that... And if one customer accesses ... |
And the overall strategy with Google's pricing is more beneficial to our end, because we basically only pay usage, and we are not feature-locked. And with the Cloudflares, Fastlys, and everybody, you basically are always feature-locked until you get to their enterprise offering. And at that point - yeah, the cost is qu... |
**Gerhard Lazu:** I was reading something about this as well, enterprise features, feature locking, things like that. It was your blog, where he said you charged for requests, right? |
**Florian Forster:** \[laughs\] Yes. |
**Gerhard Lazu:** See, I have done a bit of research; not too much, but I did notice that, where you mentioned that... And to me, that is very reasonable. You don't have to upgrade to higher price tiers just to unlock certain features. I mean, why? Does it cost you more? I mean in development time sure, but you don't g... |
**Florian Forster:** We thought long and hard about our pricing, so many times in the past... We even had like a feature locked model, closely to what Cloudflare does... And what does not reflect well in the security area is if you want to provide your customers with a security service, you should give them the means t... |
And the second thing there is like if you price by identities, customers will stop creating identities at one point, if they can choose. And we wanted to remove that sensation by telling them, "Hey, store as many things as you like, do as many things as you like. The only thing we want to have from you in return is we ... |
\[42:20\] That's really what it boils down to... Because it feels like a nice trade-off, even though - and I can be honest on that, and it's during sales meetings - it can sometimes be a problem or impediment, because people still think in users. "I want to have like a million users. I want to have a price for a millio... |
**Gerhard Lazu:** Yeah, for sure. For sure. Yeah, I mean, to me, that sounds a more sensible approach, a more honest approach, a more open approach. Everything is out there for you to use. There's like one requirement that we are able to support all your requests, and we are able to give you the quality of service that... |
**Florian Forster:** Now, looking at the date, it's like seven. |
**Gerhard Lazu:** Seven months. Okay. |
**Florian Forster:** Yeah. The thing we call Citadel Cloud now is now seven months in age... But we had a service we called Citadel v1, with kind of a different sensation to it, with the old pricing I just mentioned... And that was started in mid-2021. But we learned so many things across that journey that we needed to... |
**Gerhard Lazu:** Okay. What are the things that worked well for the cloud service in the last seven months? Good decisions, that proved to be good in practice? |
**Florian Forster:** I think it's not only directly the cloud service, but the overall change in our messaging, what we actually want to sell, and why we recommend that you use our cloud service - that message is being picked up better since like the seven last months. So that's a thing we constantly improved. So many ... |
**Gerhard Lazu:** "Upgrades? Again?! Oh, my goodness..." When did you last upgrade your phone? Serious question. |
**Florian Forster:** My phone... I'm quite pedantic, so I will catch up on releases in one to two days. \[laughter\] |
**Gerhard Lazu:** \[46:14\] Okay, that's a great answer... But for me, the updates just happen, right? I mean, unless it's like a major update, your applications on your phone - they just update. It's not a problem that people think about, or should think about. So if you run it yourself, guess what? You have to think ... |
**Florian Forster:** It's hands-off. We really call it hands-off. I mean, we take care of the TLS stuff, we will take care of updates, of backups, of rate limits, of malicious traffic... Everything is just handled for you, and that's a value I think is going great with the community... Even though Citadel's open source... |
**Gerhard Lazu:** Yeah. Yeah, that makes sense. I forgot, you're a security, right? \[laughter\] And security has this very important requirement. No, sorry, I have to run this. I mean, I understand... I can pay you to run it for me, but it has to be in the specific locations, with these restrictions, and... Yeah, that... |
**Florian Forster:** Yeah. So that's the thing I think went great with the cloud service. So free tier is definitely a thing we will reiterate on, even though you could run it on your own. I mean, it's not like there is like a feature gap, or something like that. |
**Gerhard Lazu:** Okay. What about the things that you wish you knew, before we \[48:29\]? \[laughter\] The "Oh, f\*\*k!" moments. And that will be bleeped, but... |
**Florian Forster:** \[laughs\] Yeah, there are many. Honestly, there are so many. I'm going to choose to focus on the operation side of things for the first moment. It's really like, don't try and build many funny things, even though there are great open source tools around, and everything is ready. Just try and relax... |
\[50:10\] While reflecting back, I feel like that would have helped us decrease some of the drag along operation efforts. And as well, don't run Cockroach on your own; just use the cloud service. Why not? You must really have valid and specific requirements that do not allow you to do that before you make that decision... |
**Gerhard Lazu:** Okay. |
**Florian Forster:** Other things? Yeah, our startup lesson is like "Don't assume things. Always validate things. Talk to your customers, talk to potential customers whether a feature is really needed." And we built, for example, too many features in Citadel. We assumed too many things in the past, and so we now strip ... |
**Gerhard Lazu:** Yeah. It's a good thing that you are pruning some of this stuff. Because what usually happens - you never touch it. You add it, but then you never touch it again. And that too often ends up in some very big messes, that no one wants to touch, ever... And then things just die like that, you know? |
**Florian Forster:** I mean, as soon as you see somebody raising a concern, or a bug over something, and you think, "Is that feature even being used?" And you can't really validate to yourself, "Yes, that's actively being used." You should invoke a discussion if you want to rather remove it, because just one person is ... |
**Gerhard Lazu:** I'm wondering, on the operational side of your cloud service, how do you know when something doesn't work as expected? How do you know that there is a problem, basically, is what I'm asking? |
**Florian Forster:** Let's say I am a huge fan of using existing data to get a sensation whether something's healthy or not... By which I mean we try and avoid active health checks; like, we don't ping Citadel from the outside world and figure out whether a service is available or not. I'd rather use the logfiles and t... |
And the other hugely important thing to us is like having traces, like open tracing. We use Google Cloud Tracing for that stuff, to get a sensation on whether releases change things, or smallish changes, or A/B tests also... Because sometimes we create a new branch in our open source repository, and refactor some of th... |
So it really comes down to observability based on the data you have around, but not checking it actively, because that gives you a wrong sensation. Because you're always checking happy paths there. I mean -- |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.