Merch

As a celebration of one year with Castro, we are opening a merch store. We tried a few stores and this was the best combination of quality products and a good selection. As always, all proceeds directly fund development of the app. Do not feel pressured to buy anything, I’d much rather you buy an annual subscription of Castro for a friend than a T-shirt, but we do get occasional requests for merch or about other ways to support us, so it’s there if you want it. Email support if you have any questions or issues with the site.

Castro After 1 Year

Overall I’m happy with Castro’s improvements over the past year, but we need to move iOS product out the door faster. That’s my number one focus and hopefully you will start to see those efforts bear fruit soon. Until then, I thought I’d share some of the work we’ve done that isn’t as easy for you to see but does matter a great deal.

Some might have expected a reflection on the state of the podcast industry in 2025 or Castro's place in the universe, but we have work to do. This is a post about results and what your subscription dollars are buying. Faster feed updates and a more reliable, performant podcast app.

Crash Rate

A year ago, the app crashed 1 out of every 58 sessions. Today it’s more like 1 out of every 750. We still want to improve this but the number is quite reasonable, going from the worst performing quartile according to iOS benchmarks to the second best. I obviously won’t be satisfied until we’re in the top bracket. If you expand out the chart above, it’s really the best Castro’s crash rate has ever been aside from a brief period in 2018.

If you’re wondering what we actually fixed here, there was no magic bullet.

• We rewrote CarPlay, as the old version built on MPContentItem seemed to have more and more issues on newer devices
• We rearchitected how Castro’s artwork caching works and removed some redundancies, allowing us to support episode artwork and lots of artwork in explore while improving performance and reliability
• We are building out a new data layer built on structured concurrency as well as a new networking layer, and this bypasses various threading issues and data races you might have seen in the old code
• We stayed on top of issues that crop up when we link to newer versions of iOS, so the number continues to go down instead of creeping back up

Also I’m obligated to say, as someone with a background in mobile app reliability, sessions/crash is not really the best way to parse crash data, but it’s the number Apple gives us and the fairest comparison across apps.

Outages

It was a touchy subject in the past, but there have been no sustained service outages since we took ownership. There was really only one unplanned outage of any consequence, when our cloud provider spontaneously rebooted our database server and no team members were around. I believe it lasted an hour or so before everything came back. Many apologies for that.

Nobody running a web service can tell you it will never go down, but I feel good about our track record. We will continue to do our best to avoid issues and communicate proactively when they occur. I feel especially good about this in light of the number of changes we have made to the infrastructure, enumerated below. It’s an issue I have taken incredibly seriously and I really intend for Castro to be associated with excellent service and reliability, not outages.

Backend Updates

I've compiled a non-exhaustive list of features and upgrades added to Castro’s backend, as my focus really has been on creating a sustainable infrastructure that will be around for a long time and allow us to iterate quickly and safely.

One thing to call out if you’re going to skip the list below, is that this weekend we once again dialed up our feed updates. If Castro’s feed updates aren’t already the fastest in the industry, they will be soon. Also, if you miss an episode on Castro today, it’s almost certain that the backend updated just fine but the event processing failed on the client. This is an issue I have been very focused on and expect to eliminate entirely in the next iOS release.

The List:

• Updated to Postgres 16 from 11 and moved to a more performant database server
  • Likely our most impactful fix of the year, leading to a dramatic improvement in pretty much everything
• All backend servers redeployed and updated to new images with Debian 12. Some of these had not been touched in several years.
• Updated to ElasticSearch 7 and built out a new search cluster, we are now indexing episodes and will soon be indexing transcripts
• Larger, more performant Redis servers, updated Redis
• Updated to Rails 7.2 from (5.2?)
  • We’re also using the latest version of Sidekiq Pro and other relevant dependencies
• Ruby upgraded from 2.5 to 3.3, including YJIT support
  • Of particular note was MALLOC_ARENA_MAX: 2, which was an incredible improvement in worker efficiency
• Rolled out self-serve ads
• Faster feed updates
  • A big issue Castro had in the past was that unpopular* feeds could fall hours behind, because they’d be prioritized behind updates to popular feeds and effectively be starved out. Sometimes, every popular feed would update dozens of times before the unpopular ones.
  • This now works more like a FIFO queue and Castro’s feed update mechanisms are responsive to load, so when the server gets behind, we start updating every active* feed slightly less often but still in order.
  • Just this past weekend we again cut the time between feed updates. (All numbers in this post are from before that improvement, so things are actually better than written here)
  • It’s a little hard to quantify performance before acquisition, because performance was obviously degraded in January of 2024, but *best* case scenario for an unpopular feed was updating every hour when we took over.
  • Now, the *worst* case scenario is more like a half hour but 90%+ of the time every active feed on Castro is updated at least every 8-10 minutes, often faster, with popular feeds usually within 4-5 minutes.
  • From 43 days to 2 hours to 8 minutes
  • Private and popular feeds update have seen marked improvements, but unpopular feeds were the real issue before and that issue is entirely gone.
  • **For purposes of terminology above, “active” means someone on Castro subscribes to the feed, “unpopular” feeds are active but only have a handful of listeners, the exact number has changed several times and will likely be tweaked more.
  • More info on one of many feed improvements
• Brand new explore data and APIs, brand new charts system, category charts, episode charts, etc
  • Automated locale support in charts, good explore content in every country, added many locales, etc
• podcast:transcript support, parser supporting all 4 major transcripts, as well as XML transcripts and more
• Enriched podcast data
  • Updates to podcast model
    • language / locale support, canonical links, etc
  • Updates to episode model
    • Added episode artwork (showing in app)
    • Added episode and season data (showing in app)
    • Added individual episode author (not showing in app quite yet)
  • Normalizing iTunes categories across feeds
• Better Canonicalizing
  • Added support for canonical self links, itunes_new_feed_url, etc.
  • Rebuilt feed canonicalization from the ground up incorporating redirects, itunes, self links, Uri normalization and more.
  • We’re seeing fewer dead search results, fewer duplicates in search, and by combining feeds we save on infrastructure costs
  • There is always more to do here but really happy with the strides we’ve made
• Cutover to profiles
  • With the old account system, many things were ranked according to how many accounts had ever subscribed to them. This made the power laws of podcasts even worse than they actually are, skewing search results and charts. Since we’ve cut over to profiles, you’ll notice this effect is less pronounced and the search results and charts are more reasonable.
  • Profiles allow access to subscriptions across devices, something you will hear more about soon
  • I'm very proud we were able to do this with no disruption in service.
• WebSub support
  • 17% of active feeds have updated at least once using WebSub, making your feed updates instant and saving extra work on our infrastructure
• Improvements around identifying private feeds

All this stuff adds up. Event queries, our most common endpoint by far, had a mean-90 of 40-60ms in the past, right now that number is < 15ms and literally getting better on a weekly basis. Many of our endpoints have seen even more substantial improvements.

Support

One more thing to call out is our email support is really good. An actual person with deep knowledge of the app reads every email, helping where we can and telling you when we cannot. We try to answer everything, but if we haven’t gotten back to you ping us again in a few days and we’ll get there. We also do our best to follow up again once we’ve fixed something, even if it’s months down the road.

Thanks

Thanks for reading and for using Castro. If you like what we’re doing, please buy a subscription and tell your friends or social media followers. The only way small developers who are doing things the right way can survive is if people like you continue to support them.