Re: Turning on savehist-mode by default

[Stefan Kangas <stefankangas@gmail.com>]

Adam Porter <adam@alphapapa.net> writes:

> In addition to what Po said, I'd like to gently reiterate what I said
> earlier in this thread: Given my experience with savehist causing
> unexpected and hard-to-debug performance problems[0], I'd guess that
> there are more such cases in the wild waiting to be triggered.  If, e.g.
> Emacs 30.1 enabled it by default, I can imagine a number of users
> suddenly encountering weird pauses, and they'd probably blame GC
> initially[1].
>
> As well, I have some concerns about savehist's having the potential to
> cause weird bugs in other libraries: The savehist-save function seems to
> comment out individual elements of savehist-minibuffer-history-variables
> that it determines are unreadable.  That's understandable from its
> perspective, but what effect will that have on libraries that may not be
> expecting for their data structures to have certain parts disappear
> after restarting Emacs?  I can just imagine the bug reports from users
> showing apparently corrupted or elided data structures, and having no
> clue as to what is mutating them, because the code isn't within the
> library having a bug reported against it.
>
> IMHO, a feature like this is indeed expected and overdue in Emacs, but I
> don't think savehist is a good way to enable it, given its current
> implementation.  ISTM that something built on top of the new
> multisession library would be more appropriate and (at least, have the
> potential to be) more robust.

Thanks, this is clear and compelling.

It seems to me that savehist-mode is *not* ready for being enabled by
default.  I think the best we can do at this point is to start working
on improve and/or replace it.

Would anyone like to volunteer to do that?

> 0: https://github.com/alphapapa/ement.el/issues/216

Did you think about how issues like the above could be solved (in
general, not necessarily in savehist specifically)?

> 1: Which may, in fact, be part of the problem, considering how
> savehist-save works: it prints a Lisp object to a buffer, then tries to
> read it to determine whether it's readable--doesn't that generate
> garbage?  Now consider doing that every 1-5 minutes and upon exit.

savehist (or it's replacement) could also be smarter about it, for
example you could skip re-serializing history variables that have not
been touched.