I watched Aaron Patterson's keynote from RailsConf 2020 (these talks all just came out this week, Couch Edition) and I'm struck when reading your comment that it's absolutely possible to do this, and ultimately you may or may not reach a state where it's possible to "end-of-life" one or the other alternative without making anyone cry about it.
He spends most of the second half of the talk explaining (with a profiler) how two query implementations on ActiveRecord which should be doing pretty much about the same thing, actually perform wildly differently, and narrowing it down with data to understand why.
The talk is really interesting and I think it's a parallel to the current discussion in a way because, in Aaron's talk there were two implementations and one was obviously better performing (but unfortunately it was the one that is harder to write.) He argued mostly that, unless there is a reason for these two implementations to behave and perform differently, the one that is slower (but easier to write) should become more like the one that is "enhanced with a performance pro-tip" that arguably the ActiveRecord user shouldn't really need to know about. In other words, that the bad performance was actually a bug.
(tl;dr: the pro-tip is, to build your own pre-sanitized SQL WHERE clause strings, because you may save something like 38% performance in a pathological case, recovered time that the SQL engine was going to spend traversing AST and compiling.)
My point is that, exactly. If there is bad performance over here, it is a bug. Bad performance over there, bug. But bad performance over here, that can only be solved by swapping for bad performance over there?
This does not sound like a bug anymore, it starts to sound more decidedly like a trade-off. But given enough time, perhaps they will fix the bugs, and the two can be made to perform similarly, fixing bugs on either side, and the need for two alternatives will go away. So why would that be bad?
(And if it can't be done, or can't be done yet, why wouldn't we expect to see both implementations maintained as long as needed to find a solution that pareto-dominates them both?)
Sure. But I don't completely see the point. "bugs" you don't own are the worst bugs because you have no say on when or if they are getting fixed. I deal with such things on a fairly regular basis - our RDBMS (SQL Server) acting in pathological ways and when I look into it, people have been complaining about this stuff for a decade.
Ultimately, the software creator determines what is and isn't a bug - formally through what they label them as, but effectively through how they act upon the behavior. As an end user of a system, things like the above are less like bugs, and more like features.
> But given enough time, perhaps they will fix the bugs, and the two can be made to perform similarly, fixing bugs on either side, and the need for two alternatives will go away. So why would that be bad?
Because (possibly unlike with your Rails example?) the suggestion that this is possible to do by improving the VM side is, to be blunt, a lie. A VM is a fundamentally leaky abstraction, in far more ways than just the timing behavior (i.e. performance), and the need for both won't go away by just "fixing" the VM side. It's literally impossible for them to ever make a VM "feel the same" as WSL1. To give you a taste: My "feel" was that I wouldn't have to deal with all the VHD/disk/partition/volume management I was so happy to finally get away from. My "feel" for WSL1 was that I could compress its files with the WoF LZX compression in Windows 10 to save a ton of disk space and yet still read them with native Windows tools as if nothing was different about them. My "feel" for WSL1 was that I could actually read and even move Linux files around directly from inside Windows as long as WSL1 wasn't running, treating a WSL folder just like a regular folder. My "feel" was that I could even use Windows tools to alter these files as long as I maintained the Linux attributes correctly.
And my "feel" was that using WSL would have zero impact on any VM technologies that I may or may not want to use, but that's another can of worms.
I could go on, but hopefully you get the point. Either you have to dismiss my "feel" for being somehow "invalid" to your liking, or you have to realize you'll never get the same feel.
> My point is that, exactly. If there is bad performance over here, it is a bug. Bad performance over there, bug. But bad performance over here, that can only be solved by swapping for bad performance over there? This does not sound like a bug anymore, it starts to sound more decidedly like a trade-off.
Only if you assume "only solve X via Y" is correct. I don't believe that's true. I believe they just haven't put in the required work yet, and that if they did put in the work, they could still improve WSL1 substantially. The real reason for why they haven't done so is anyone's guess, but to me, there are multiple signs (not just I/O-performance-related) that suggest the WSL team has to do most of their work independently and can't necessarily get sufficient cooperation from the teams working on different subsystems, and hence this is why they've been unable to improve things. Or maybe the folks who can work on the required parts are no longer available. Regardless of the reason though, I don't believe they've reached the limits of the technology yet.
> Because (possibly unlike with your Rails example?) the suggestion that this is possible to do by improving the VM side is, to be blunt, a lie. A VM is a fundamentally leaky abstraction, in far more ways than just the timing behavior (i.e. performance), and they need for both won't go away by just "fixing" the VM side.
The fun part about my example is actually that he saw a 60% performance difference, but only recovered 38% of that by fixing the one bug that was exposed in the talk.
It was decidedly a bug. Refusing to cop out and say "it's a necessary evil, and can't be fixed" got him that 38% boost in one shot. The remaining ~22% is still a problem that has yet to be explained completely, but now it's not as big (or soon won't be, when they resolve the issues created by fixing the bug, which AIUI it couldn't be merged yet...)
I do get your point, and I grok that parts of this problem simply can't be fixed outright. But from my perspective, the bug isn't that WSL2 is missing features of WSL1, it's that WSL1 performs 60% slower for my use cases (*not an actual measurement) while doing what appears to be, at least from my perspective, basically what is meant to be the same task. I don't really know why, except by what some greater experts have shared, some of that which I have acquired by osmosis.
Your use cases are valid, and we mostly agree, when you say the limits of the technology have not been reached, it seems like we're both arguing that there are simply more bugs yet.
He spends most of the second half of the talk explaining (with a profiler) how two query implementations on ActiveRecord which should be doing pretty much about the same thing, actually perform wildly differently, and narrowing it down with data to understand why.
The talk is really interesting and I think it's a parallel to the current discussion in a way because, in Aaron's talk there were two implementations and one was obviously better performing (but unfortunately it was the one that is harder to write.) He argued mostly that, unless there is a reason for these two implementations to behave and perform differently, the one that is slower (but easier to write) should become more like the one that is "enhanced with a performance pro-tip" that arguably the ActiveRecord user shouldn't really need to know about. In other words, that the bad performance was actually a bug.
(tl;dr: the pro-tip is, to build your own pre-sanitized SQL WHERE clause strings, because you may save something like 38% performance in a pathological case, recovered time that the SQL engine was going to spend traversing AST and compiling.)
My point is that, exactly. If there is bad performance over here, it is a bug. Bad performance over there, bug. But bad performance over here, that can only be solved by swapping for bad performance over there?
This does not sound like a bug anymore, it starts to sound more decidedly like a trade-off. But given enough time, perhaps they will fix the bugs, and the two can be made to perform similarly, fixing bugs on either side, and the need for two alternatives will go away. So why would that be bad?
(And if it can't be done, or can't be done yet, why wouldn't we expect to see both implementations maintained as long as needed to find a solution that pareto-dominates them both?)