-
-
Notifications
You must be signed in to change notification settings - Fork 384
engineering practices
Tyler Neely edited this page Nov 15, 2019
·
1 revision
Over the years that sled development has been active, some practices have been collected that have helped to reduce risks throughout the codebase.
- Start with the correctness requirements, ignore the performance impact until the end. You'll usually write something faster by focusing on keeping things minimal anyway.
- Throw away what can't be done in a day of coding. when you rewrite it tomorrow, it will be simpler.
- Don't do what can't be tested to be correct
- For concurrent code, it must be delayable to induce strange histories when running under test
- For IO code, it must have a failpoint so that IO errors can be injected during testing, as most bugs in cloud systems happen in the untested error-handling code
- Lean heavily into model-based property testing. sled should act like a BTreeMap, even after crashes
- cold code is buggy code
- if you see a significant optimization that will make correctness-critical codepaths harder to hit in tests, the optimization should only be created if it's possible to artificially increase the chances of hitting the codepath in test. Fox example, sled defaults to having an 8mb write buffer, but during tests we often turn it down to 512 bytes so that we can really abuse the correctness-critical aspects of its behavior.
- No silent truncation should ever occur when converting numbers
- No silent wrapping should occur
- Crash or return a ReportableBug error in these cases
-
as
is forbidden for anything that could lose information - Clippy's cast lints help us here, and it has been added to all pull requests
- dependencies should be minimized to keep compilation simple
- Self should be avoided. We have a lot of code, and it provides no context if people are jumping around a lot. Redundancy here improves orientation.