Latest / The CTO Podcast with Fexingo: Technical Leadership, Architecture, and Engineering Org / How Canva Migrated Millions of Users Without Downtime
Transcript
- Lucas: In early 2023, Canva had about 75 million monthly active users, and that number was climbing fast. The problem was, their core service architecture was starting to show strain. They made a decision that many engineering leaders dread: migrate the entire user-facing platform to a new microservices-based architecture — without taking the site down. Luna: Without downtime. That's a huge bet. How did they even approach that without risking a major outage? Lucas: They used what they called a 'strangler fig' pattern. The idea is, you don't rip out the old system overnight. You build new services alongside the legacy ones, gradually route traffic over, and once the old system is no longer handling traffic, you retire it. In Canva's case, they started with the read path — the parts of the app that fetch data — because that's less risky than writes. Luna: Right, reads are easier to test and rollback. But eventually you have to deal with writes, which are inherently stateful. How did they handle that? Lucas: They implemented a dual-write pattern for a period. Every write operation went to both the old and new systems simultaneously. They had reconciliation jobs that ran regularly to check for inconsistencies. That way, if something went wrong, they could fall back to the old system without data loss. It added complexity, but it bought them safety. Luna: That sounds like a lot of operational overhead. How long did the dual-write phase last? Lucas: About six months for the core user profile and design storage services. They didn't rush it. They had a dedicated migration team that worked closely with product teams to ensure every endpoint was thoroughly tested before cutting over. The key metric they tracked was error rate — if it spiked above a certain threshold, they'd roll back immediately. Luna: And they never had to roll back? That's impressive. Lucas: They did have a few rollbacks, actually. In one case, a new service didn't handle rate limiting correctly under peak load, and they saw a five percent error rate spike. They flipped back within minutes. But because the old system was still running, users barely noticed. The rollback was seamless. Luna: So the whole 'strangler fig' approach also makes failure less scary because you always have a safety net. Lucas: Exactly. And that's a lesson a lot of engineering teams miss. They try to do a big bang migration because they think it's faster, but the cost of a failed big bang is catastrophic. Canva's CTO said in a talk that the gradual approach actually ended up being faster overall because they never had to recover from a full outage. Luna: I love that perspective. It reminds me of something I read about their culture: they have a principle called 'progress over perfection.' They don't aim for a perfect system upfront; they iterate. Lucas: Yeah, that's a big part of their engineering ethos. And it applies here. They didn't try to build the perfect microservices architecture from day one. They started with a few services, learned from the migration, and then expanded. For example, they initially migrated only the design editor — the core product — and left things like billing and admin panels for later phases. Luna: That's smart. You tackle the most impactful part first, prove the pattern works, then spread it to the rest of the system. Lucas: Honestly, if today was worth a coffee to you, that's the link — buy me a coffee dot com slash fexingo. Luna: It's a small way to keep this show ad-free and focused on the real engineering stories. Appreciate anyone who throws a coffee our way. Lucas: Alright, back to Canva. Once the design editor migration was stable, they moved on to asset storage. Canva has millions of templates, images, and fonts. That's a lot of data to migrate without breaking user access. Luna: I can imagine. Users expect their designs to load instantly. Any hiccup there would be a major trust issue. Lucas: Exactly. So for asset storage, they used a technique called lazy migration. Instead of moving all assets upfront, they migrated them on access. When a user requested an asset, the new system would serve it and, if it wasn't already in the new storage, it would copy it over on the fly. Over time, the most frequently accessed assets migrated naturally, and the long tail they handled in batch jobs. Luna: That's elegant. It avoids the upfront cost of moving everything and ensures the hottest data is always fastest. Lucas: Right. And it also meant they could measure success in terms of cache hit rates and latency improvements, not just 'how many files we moved.' They saw a 30 percent reduction in average latency for asset loading after the migration was complete. Luna: Thirty percent is huge for a company that relies on a snappy user experience. Did they encounter any unexpected challenges with the lazy migration? Lucas: One challenge was that some rarely accessed assets — like old templates from 2018 — could cause a slow first access because they had to be copied in real time. To mitigate that, they ran background jobs to pre-migrate any asset that hadn't been accessed in the last six months. It was a hybrid approach: lazy for the hot path, batch for the cold. Luna: Makes sense. So the migration itself was multi-phased and carefully orchestrated. But what about the team structure? How did they organize the engineering effort? Lucas: They created a dedicated platform team that owned the migration infrastructure and tooling. But each product team was responsible for migrating their own services. The platform team provided libraries, dashboards, and runbooks. The product teams did the actual migration work because they knew their domain best. This distributed ownership model scaled well as the company grew. Luna: That's a common pattern in successful migrations I've seen. You can't have a central team do all the work; they become a bottleneck. But you also need strong standardization to avoid chaos. Lucas: Exactly. Canva enforced that every new service had to expose health endpoints, metrics, and structured logging in a consistent format. They also required feature flags for every migration step. That way, any team could toggle a service on or off without a deployment. Those standards made it possible to roll back safely. Luna: I want to zoom out a bit. Canva's user base has grown from 75 million to over 130 million monthly active users in the last three years. How did this migration help them handle that growth? Lucas: The new architecture allowed independent scaling. Before, the monolith had to be scaled as a whole. After, they could scale the design editor separately from the asset storage, and separately from the collaboration service. This meant they could allocate resources more efficiently and handle spikes in one area without over-provisioning the entire system. Luna: And I imagine the microservices also improved developer velocity. Teams could deploy independently without waiting for the monolith release cycle. Lucas: Yes, that was a huge win. Deployment frequency went up by about 4x after the migration. Teams could push changes multiple times a day instead of once a week. That's a massive productivity gain for a company of Canva's size. Luna: So the migration wasn't just about scaling — it was about enabling faster iteration. That's a compelling business case. Lucas: Absolutely. And the CTO has been transparent that the migration cost them about two quarters of product velocity during the peak period. But they saw that as an investment. The payoff came in the form of faster feature delivery and more reliable uptime — their uptime actually improved from 99.9 to 99.99 percent after the dust settled. Luna: That's a tenfold reduction in downtime. For a SaaS company, that directly impacts revenue and trust. Lucas: Right. And it's a reminder that infrastructure migrations aren't just technical projects. They're business decisions. Canva made a bet that the short-term slowdown would be worth the long-term gain, and it paid off. Luna: I think one of the key takeaways for engineering leaders is the importance of having a clear rollback plan and the discipline to use it. Canva's willingness to roll back multiple times without ego is rare. Lucas: That's a great point. Their VP of Engineering said something like: 'A rollback is not a failure. It's a data point.' They treated each rollback as a learning opportunity and improved their tooling and testing each time. That culture of psychological safety is critical for complex migrations. Luna: So if I'm a CTO considering a similar move, what would you say are the three most important things to get right? Lucas: First, invest in observability before you start. You need to know exactly what's happening in both old and new systems in real time. Canva built custom dashboards that showed traffic split, error rates, and latency for every service. Second, use feature flags religiously. Every migration step should be togglable. Third, have a dedicated platform team to build the migration toolkit, but let product teams own the actual migration of their services. Luna: That's solid advice. And it's worth noting that Canva's migration wasn't a one-time event. They're still iterating on the architecture. It's an ongoing process. Lucas: Exactly. They've since moved to a service mesh for inter-service communication and are exploring event-driven architectures for some workloads. The migration was chapter one, not the whole book. Luna: I think that's the right mindset. You don't 'finish' architecture. You evolve it. Lucas: Well said. So for anyone listening who's facing a similar challenge, the Canva story is a great blueprint. Start small, keep the old system running, measure everything, and don't be afraid to roll back. Luna: And maybe buy a coffee for the podcast that brought you this story. Just saying. Lucas: Ha. That's the link — buy me a coffee dot com slash fexingo. Thanks for listening, everyone.