#+TITLE: el-job #+AUTHOR: Martin Edström #+EMAIL: meedstrom@runbox.eu #+EXPORT_FILE_NAME: el-job #+TEXINFO_DIR_TITLE: El-Job: (el-job). #+TEXINFO_DIR_DESC: Async multicore mapcar #+TEXINFO_DIR_CATEGORY: Emacs #+HTML: Packaging status Imagine you have a function you'd like to run on a long list of inputs. You could run =(mapcar #'FN INPUTS)=, but that hangs Emacs until done. This library lets you run the same function in many subprocesses (one per CPU core), each with their own split of the =INPUTS= list, then merge their outputs and pass it back to the current Emacs. In the meantime, current Emacs does not hang at all. Best of all, it completes /faster/ than =(mapcar #'FN INPUTS)=, owing to the use of all CPU cores! That's it in a nutshell. You can look at real-world usage by searching for "el-job" in these packages: - [[https://raw.githubusercontent.com/meedstrom/org-mem/refs/heads/main/org-mem.el][org-mem.el]] - [[https://raw.githubusercontent.com/meedstrom/org-roam-async/refs/heads/main/org-roam-async.el][org-roam-async.el]] * 2.7.0 - New entry point: =el-job-parallel-mapcar=! Read more below. - Argument =:inputs= are now split differently, for the sake of a predictable order of outputs. It is likely less optimal, but hopefully not too much. - Argument =:require= now accepts strings in addition to symbols. Strings are passed to =load= instead of =require=. - Value of =temporary-file-directory= no longer added to =:inject-vars= for you. - Now the only values added are =load-path= and =native-comp-eln-load-path=. - Function =el-job-ng-job= renamed to =el-job-ng-get-job=. - New class =el-job-ng-job=. - Delete many aliases to el-job-old.el. ** New entry point: =el-job-parallel-mapcar= Until now, we had to make do with an unwieldy =el-job-ng-run= with ~5 keyword arguments. And we had to understand asynchronous programming to use it. I've long dreamed to be able to brainlessly rewrite a form =(mapcar #'FN INPUTS)= to something like =(multicore-mapcar #'FN INPUTS)= and have it Just Work and Just Be Faster. Finally, it is done! The calling convention is : (el-job-parallel-mapcar FN INPUTS &optional INJECT-VARS) Please see the docstring for what you need to know. *** Weaknesses Do not use =el-job-parallel-mapcar= with overly trivial functions. It adds some overhead per item, so it actually slows you down if the function itself only takes microseconds or nanoseconds per invocation. Example of a bad use-case: #+BEGIN_SRC elisp (let ((spam (make-list 1000000 "spam"))) (list (benchmark-elapse (mapcar #'upcase spam)) (benchmark-elapse (el-job-parallel-mapcar #'upcase spam)))) #+END_SRC Return value: : (0.482017027 4.793430311) Another bad use-case is when plain =mapcar= would've been fast enough. There is some constant overhead related to spinning up subprocesses. Changing 1000000 from the earlier expression to just 100 reveals the constant to be ~210ms on my machine: #+BEGIN_SRC elisp (let ((spam (make-list 100 "spam"))) (list (benchmark-elapse (mapcar #'upcase spam)) (benchmark-elapse (el-job-parallel-mapcar #'upcase spam)))) #+END_SRC Return value: : (0.000041974 0.219570936) * 2.6.0 - For =el-job-old-launch=, new argument =:eval= * 2.5.0 Released [2025-10-06 Mon], v2.5.0 comes with a variant library "el-job-ng". I find it simpler and easier to reason about. 400 lines of code instead of 800. Some differences: - Does /not/ ever keep a process alive - Does /not/ merge the subprocesses' outputs in a bespoke way, just uses =append=. - Does /not/ look up =load-history= to try hard to find an .eln variant of your libraries, instead it lets =require= look up =load-path= normally. - Removed argument =:if-busy=, you can manage this yourself with =el-job-ng-busy-p= and =el-job-ng-kill=. - Argument =:funcall-per-input= now takes a function of two (2) arguments, not one. - Argument =:callback= now takes a function of one (1) argument, not two. - Added optional argument =:eval=. The design is always changing in response to my needs, so feel free to file an issue or email! It's instructive to hear about other people's needs. ** Future work I may write yet another variant. Something that came with experience is that it's best to make a new variant library for a narrow use-case, rather than complicate one library with different code flows. When it comes to this type of library, you really want to keep it easy to reason about! Ideas as of [2025-10-06 Mon]: - File IPC :: Sending input and output by writing and reading files, instead of through the pipe connection. Theory: performance can sometimes be a lot better, with large inputs and outputs. I would guess it depends a lot on the machine. Drawback: if the data sent is sensitive, those files probably should be encrypted, and that could negate any performance benefit. - Worker daemons :: Keeping subprocesses alive forever, so they are available at beck and call -- think worker daemons. That's basically implemented in "el-job-old", and partly why it got so hairy, but it could be remade to put this usage front-and-center, with a "happy path" UX. At this time, I do not have a demanding use-case to experiment with, to discover what that happy path should be, how the affordances should be. If you need this now, I recommend [[https://github.com/jwiegley/emacs-async][async.el]]. * README for 2.4.8 ** Design rationale I wanted to shorten the round-trip as much as possible, *between the start of an async task and having the results*. For example, say you have some lisp that collects completion candidates, and you want to run it asynchronously because the lisp you wrote isn't always fast enough to avoid the user's notice, but you'd still like it to return as soon as possible. *** Processes stay alive In the above example, a user might only delay a fraction of a second between opening the minibuffer and beginning to type, so there's scant room for overhead like spinning up subprocesses that load a bunch of libraries before getting to work. Thus, el-job keeps idle subprocesses for up to 30 seconds after a job finishes, awaiting more input. An aesthetic drawback is cluttering your task manager with many processes named "emacs". Users who tend to run system commands such as =pkill emacs= may find that the command occasionally "does not work", because it actually killed an el-job subprocess, instead of the Emacs they see on screen. *** Emacs 30 =fast-read-process-output= Some other libraries, like the popular [[https://github.com/jwiegley/emacs-async/][async.el]], are designed around a custom process filter. Since Emacs 30, it's a good idea to instead use the /built-in/ process filter when performance is critical, and el-job does so. Quoting [[https://github.com/emacs-mirror/emacs/blob/master/etc/NEWS.30][NEWS.30]]: #+begin_src org ,** The default process filter was rewritten in native code. The round-trip through the Lisp function 'internal-default-process-filter' is skipped when the process filter is the default one. It is reimplemented in native code, reducing GC churn. To undo this change, set 'fast-read-process-output' to nil. #+end_src ** News 2.4 - Jobs must now have =:inputs=. If =:inputs= nil and there was nothing queued, =el-job-launch= will no-op and return the symbol =inputs-were-empty=. ** News 2.3 - Some renames to follow Elisp convention - =el-job:timestamps= and friends now =el-job-timestamps=. ** News 2.1 - DROP SUPPORT Emacs 28 - It likely has not been working for a while anyway. Maybe works on the [[https://github.com/meedstrom/el-job/tree/v0.3][v0.3 branch]], from 0.3.26+. ** News 2.0 - Jobs must now have =:id= (no more anonymous jobs). - Pruned many code paths. ** News 1.1 - Changed internals so that all builds of Emacs can be expected to perform similarly well. ** News 1.0 - No longer keeps processes alive forever. All jobs are kept alive for up to 30 seconds of disuse, then reaped. - Pruned many code paths. - Many arguments changed, and a few were removed. Consult the docstring of =el-job-launch= again. ** Limitations 1. The return value from the =:funcall-per-input= function must always be a list with a fixed length, where the elements are also lists. For example, org-mem passes =:funcall-per-input #'org-mem-parser--parse-file= to el-job, and if you look in [[https://github.com/meedstrom/org-mem/blob/main/org-mem-parser.el][org-mem-parser.el]] for the defun of =org-mem-parser--parse-file=, it always returns a list of 5 items: #+begin_src elisp (list (if missing-file (list missing-file)) ; List of 0 or 1 item (if file-mtime (list file-mtime)) ; List of 0 or 1 item found-entries ; List of many items org-node-parser--found-links ; List of many items (if problem (list problem)))) ; List of 0 or 1 item #+end_src It may look clunky to return sub-lists of only one item, but you could consider it a minor expense in exchange for simpler library code. 2. Some data types cannot be exchanged with the children: those whose printed form look like =#<...>=. For example, =#=, =#=, =#=. IIUC, this sort of data only has meaning within the current process -- so even if you could send it, it would not be usable by the recipient anyway. 3. For now, this library tends to be applicable only to a narrow set of use-cases, since you can only pass one =:inputs= list which would tend to contain a single kind of thing, e.g. it could be a list of files to visit, to be split between child processes. In many potential use-cases, you'd actually want multiple input lists and split them differently, and that's not supported yet.