Revisit checking in enough source to build "small-icu" into the node repo

[srl295]

from discussion on #3460 and others

Theoretically, a stripped down ICU source base might be 23M on disk / 4.3M compressed - this is including the 1.5M "small ICU" (i.e. English only) .dat file in lieu of the 25M one which usually comes in an ICU tarball.

Investigate the actual size needed here. This would include full functionality, just not full data. If you want node with full data you would still be able to: (a) download a full ICU tarball and build with full-icu as an option (just as you can do today) or (b) provide the full .dat file at runtime (just as you can do today)

So the upshot of this change would be that we could make small-icu the configure default without any download needed, and a small impact to the repo (probably checking in source+blob, otherwise checking in a small tarball).

cc @trevnorris @bnoordhuis @jasnell

[jasnell]

+1 ... with the caveat that we need to investigate how this impacts building on the smaller devices (like rpi). If the impact building icu by default on those devices is significant enough, then we'll need a configure option for optionally turning off the ICU build and set that as the default of CI runs only on the rpi devices.

[srl295]

@jasnell you could still use --with-intl=none and get today's configure behavior. Should be noted here as well.

[trevnorris]

I'm alright further investigating this approach. The 1.5MB dat file would only be updated every six months or so would have less impact then updating v8 does.

[jasnell]

I would add that I'd definitely prefer the source to be checked in directly rather than as a tar ball, even if it does mean the full 23 M. The tarball idea just feels dirty and reduces visibility (well, it's no less visible than what we have now, but ... you know). With changes coming only twice per year, the hit (shouldn't be / won't be) too bad.

[trevnorris]

The source sounds fine. v8 uses ~98MB and openssl uses ~43MB.

[srl295]

@silverwind asked

BTW, is building with full-icu by default out of question?

First of all, I'd certainly like that, personally.

Second, "Download-by-default" is out of the question, so to do full-icu by default probably means checking in ICU + full data- which would be something like ~45M on disk (source files) and 13M compressed (delta to the nodejs tarball). The "" has to do with how much of the tooling needs to be stripped out or left ( some might be needed to do an endian swap or to rebuild small vs full data etc.)

I basically have the action here to actually try out these options and report back the deltas.

[silverwind]

@srl295 Does having full-icu packaged have any impact on performance or memory, or is it all loaded on-demand?

[srl295]

@silverwind It's all demand paged. That's why there's a separate big and little endian version.

[silverwind]

Sounds good, let's see some deltas 😉

[MylesBorins]

is there any progress on this?

[srl295]

"du -sh" reports 45M on disk. No config or source changes needed. Small enough?

[srl295]

45M is enough Icu to build full or small.