[harrygeez]

I work in a bank and we have a team of people writing microservices on top of mainframe "programs". I'm not sure how it works in the mainframe because I work mostly on the front-end, but apparently what we are used to in modern REST apis are just called programs in the mainframe. Anyway, I digress, what we are experiencing is quite the opposite: operations written in the microservice run much faster and we are kind of bottlenecked by the mainframe.

I also recently came to learn that the mainframe run jobs in batches, so I'm not sure how does that affect performance, if it even does.

[InfiniteRand]

I think this hints at the best practice, establishing APIs around the edges of the legacy system, then sectioning off chunks of legacy functionality, building internal APIs inside the legacy code around those chunks and then replacing those chunks with modern systems and making those APIs external (external as in legacy -> modern instead of legacy -> legacy). Some legacy core pieces might never be replaced but those can be contained so the bottleneck is limited. (a good example is genuinely batch operations likely nightly email alert processes, while it might be nice to modernize those)

My limited experience with this is this strategy is that it works, although it is not error proof and can be messy. You also end up discovering that assumptions you made about the legacy software are not correct and then you need to back pedal and potentially throw away previous work (ironically your own assumptions about how to modernize the legacy code become their own legacy that you need to work around as the project ages).

It's good if you have some of the original developers still around and probably worthwhile to give them consulting $$ in order to bring them out of retirement.

Perhaps the best thing about this strategy is if you in the end conclude it isn't worthwhile, you will still have modernized some sections of the codebase. Of course, because of this you need to make sure that the individual section by section modernization efforts are themselves a net gain, instead of just hoping that the final end of the modernization process will be a benefit, because you might never get to that end point.

Also, you might be able to enhance the product early on by being able to iterate and improve the sections that have been modernized, instead of waiting till everything is done to flip a switch and get new features.

The downside is this is a slow and gradual process. In the meantime your developers are filling their heads with details of a unique system that is hopefully going to go away never to be seen again.

If you have specs and documentation and tests and all of that good stuff, it might be better to just start from scratch, and build things up according to requirements.

[jacquesm]

That's the way to go about it. It won't be cheap nor fast but you'll get the job done eventually.

[OGWhales]

This is what I see most shops doing, including ours. I think the cost of getting off a mainframe is too high, so this will likely continue to be be the trend for the foreseeable future.

Also, mainframes can run both batch jobs and "online" (CICS) jobs. Our shop is about half and half with batch and online programs. The main thing that batches can be annoying with is that they run on schedules and may have to wait for other jobs to complete before they can kick off, causing delays. We just upgraded from a z13 to z15, so our mainframe is no longer a concern but we are also much smaller than a bank.

[rbanffy]

IBM's profit margins are so high they sell you a machine that's fully configured because they know you'll pay for the upgrade that's already delivered but deactivated in a year or so.

If moving off the mainframe becomes cheaper, they have a lot of flexibility

[dralley]

You seem to be conflating "throughput" and "latency" here. Mainframes are very much focused on the former.

[leokennis]

Mainframes (Z/OS) will mainly run batches (most often COBOL programs controlled by JCL jobs) but there is also CICS (https://en.wikipedia.org/wiki/CICS) which runs COBOL (and probably others) "on demand".

My experience with connecting UNIX to Z/OS was not with a real API (HTTPS) but UNIX > TIBCO EMS > something called TIBCO Substation > CICS. It was horrible, mainly because of Substation which needs to translate on the fly between two completely different platforms.

[beh9540]

I can second this - our mainframes average response time is 2-5 seconds based on the call, with 99 percentile responses in the minutes per transaction. This is after a fortune was spent moving us to a "faster" platform.

Also, at our bank, very few applications use the REST API. Since most applications are developed by contractors, most vendors don't want to use our custom API, and instead opt to connect to the mainframe directly, because then they can reuse the code at other banks after charging us to develop it.

[dralley]

Isn't this confusing latency with throughput? Mainframes are designed for throughput, raw TXNs per second, and a lot of the architecture is designed around batching. Which is great for throughput, but the antithesis of latency.