This post is a collection of mistakes that I have made in my career of software development. I want to maintain such a list and reflect on them occasionally, so I can avoid making the same mistakes again. The list is definitely incomplete and will be updated from time to time. If there are any additional mistakes of mine that you remember, I will truly appreciate it if you could remind me in the comments (or email me).
-
I had a wrong and outdated belief in BibTeX, which was corrected by Nick Bart in the bookdown PR. He first pointed out my mistakes in the issue #754. In particular, he thought that my assessment of
pandoc-citeproc
was both “factually inaccurate and needlessly disparaging”. I admit my ignorance and really appreciate Nick’s correction.Lessons learned:
Do not judge tools that I’m not really familiar with. If I have to assess them, do more research or ask experts for input, to avoid spreading misinformation or unfair assessments. It still surprises me that the Chicago Manual of Style calls for inverting the first and last name of only the first author when there are multiple authors in a bibliography entry. This has been such a weird style to me that I thought it was a bug of
pandoc-citeproc
, but I was completely wrong. -
Cropping PDF figures in knitr requires both
pdfcrop
andghostscript
. Originally I only checked the existence ofpdfcrop
and assumed that users would read its documentation to figure out that they would also need to installghostscript
, but it turned out that my assumption was not appropriate. The problem of the missingghostscript
has confused multiple users. Five years after the knitr issue was reported, I finally realized that this problem was too difficult for average users to debug, so I made changes in knitr and rmarkdown to disable this feature by default whenghostscript
is not available.Lessons learned:
When the changes I need to make in the code are relatively simple, I should make the changes, instead of requiring users to read the documentation of a third-party software package.
-
In 2013, Jared Lander reported that the tilde was too high in the code blocks of PDF output generated from
.Rnw
documents. I was rather impatient at that time and said the rude words “This is where I absolutely hate LaTeX.” Eight years later, Berry Boessenkool reported the same issue again. This time I calmly assessed the problem, and decided to provide a fix instead of simply hating LaTeX. The fix turned out to be simple enough. LaTeX is an amazingly powerful typesetting system. To be honest, I still do not prefer authoring documents with it, because the typing effort is too high to me and I often find LaTeX commands distracting in the document body (to be equally honest, I even find the Markdown syntax distracting sometimes, especially the hyperlink syntax[text](url)
). However, I think it is unfair to say I hate LaTeX simply because of a high tilde. The tilde problem is surprising and annoying, but it is just a small annoyance that could be solved automatically. I should not deny the whole LaTeX system because of this problem, especially when solutions exist. Besides, there might be a historical reason why the tilde is too high. Its author might regret it now but it is too late. As a software engineer myself, do I have regrets in my work that are too late to change? Of course I do.Lessons learned:
Be tolerant of small annoyances in large software systems. Let him who is without sin cast the first stone. Unfortunately, I was still making this mistake in 2020 (mocking at LaTeX, Word, PowerPoint, and beamer, etc.).
-
When I started writing and promoting the knitr package in late 2011, I highlighted the improvements compared to Sweave too much, and shouted out loud my frustrations with Sweave publicly for too many times. In retrospect, I appeared to be ungrateful to Sweave and its authors (especially Friedrich Leisch). Unfortunately, it took me nearly ten years to realize how deeply hurtful this could be, after a similar thing happened to myself. I have revised some of my web pages to treat Sweave more fairly, and credit where credit is due. Now I just wish Fritz could forgive me someday. To be fair, Sweave was a quite clever and novel implementation of Literate Programming (LP), and is still probably one of the very few faithful implementations of LP (no, I still do not think IPython or Jupyter is a true implementation of LP, despite its extreme popularity).
Similarly, my wording was also too aggressive about TeX Live when promoting TinyTeX from 2017. Thankfully, Karl Berry and Norbert Preining have been forgiving to me, and Karl even invited me to write an article about TinyTeX in the TUGBoat!
Lessons learned:
Recognize the second-mover advantage. Show respect and be grateful to the pioneer in the field. As the second mover, I get the pioneer’s ideas and work for free (which is especially true for free and open source software), and the hands of the pioneer are often more tightly tied than mine, since the second mover does not have any historical debt. It is often easier to start afresh than to maintain an existing package.
-
In 2017, after supporting tinytex in the R Markdown ecosystem, I started to think about using tinytex for Rnw documents when users click on the Knit button in RStudio. Rnw documents could already take advantage of tinytex if they are compiled via
knitr::knit2pdf('test.Rnw')
, i.e.,knitr::knit2pdf()
will calltinytex::latexmk()
to compile the intermediate.tex
to.pdf
, which means missing LaTeX packages will be automatically installed. However, I suspected that most RStudio users might be used to the Knit button, which did not usetinytex::latexmk()
at that time, so I filed a feature request to the RStudio IDE. My colleague Kevin Ushey finally implemented it. Two things surprised me in this process:-
I got 17 thumb-ups in my feature request in the rstudio repo on GitHub—originally I thought few people would care about it, e.g., I did not know that Nick still uses Rnw.
-
Emmanuel Charpentier appeared to be uncomfortable with me saying “I don’t think Rnw is still widely used today,” and argued that Rnw and LaTeX were still used and indispensable.
For some reason, Emmanuel misunderstood me. I intended to help Rnw users, but he thought I was going to kill Rnw. How come my helping hand (in my eyes) became the hand of a murderer? I just could not understand the misunderstanding, until I was hurt by a “helping hand” a few years later, and the helping hand could not understand how it actually pushed me into a hell.
Lessons learned:
Be aware of asymmetric power dynamics. R Markdown is likely to be on the more powerful side now, and LaTeX is on the less powerful side in the R community. If I want to support the latter better, I should just go ahead and do the work, and should not have brought up “popularity” in the mean time. “Popularity” is a sensitive and dangerous topic to be mentioned, and worse yet, often irrelevant. Remember, influenza is also “popular,” but you probably do not want to get it because of its popularity. How do I know R Markdown is not flu? I cannot really tell.
How many Rnw users are there today? I still do not know. Since I do not have data, I’m not entitled to say “Rnw is not widely used today” or “R Markdown is more popular.” As an R Markdown developer, I’m definitely biased.
-
-
I have hoped to learn more about Jeroen’s curl package for a long time, because many web services provide APIs now; curl is rather low-level, but sometimes I’m just interested in learning low-level stuff (except for the C language). Finally I got an incentive—I received a notice from CRAN, which recommended package authors to stop using
ftp://
. I used it inxfun::upload_win_builder()
. I have been wondering if it is possible to automate the upload via https:// instead. Then I discoveredcurl::handle_setform()
andcurl::form_file()
. Problem solved.Then I became more ambitious, and wanted to tackle a problem that I had never completely understood before—calling an API with authentication. I had a vague sense that the authentication info should be included in a HTTP request header, but I did not know the format or how to add the info. It did not take me long to discover
curl::handle_setheaders()
. Since I also need to send a file in the request, I triedcurl::curl_upload()
, but this function does not support setting headers directly. It was quite straightforward to add the support, so I sent the PR jeroen/curl#243. Then Jeroen told me I could just use thehttpheader
argument, which is equivalent to usinghandle_setheaders()
. I was glad to learn this simple solution, but when I tried it, it did not work.I tried to make a minimal reprex but was not able to reproduce the problem in simpler cases. After a couple of hours of investigation, it turned out that I accidentally typed the colon in
Authorization: blabla
with my Chinese input method turned on, which gave me a full-width colon. It was hard to tell the difference betweenAuthorization:blabla
andAuthorization: blabla
. How did I discover it eventually? Well, I copied the header from the working example, and compared it with the the header in my failing code. I was taken aback by the result:> 'Authorization:blabla' == 'Authorization: blabla' [1] FALSE
I could not believe my eyes. I read every single letter, A, u, t, h, o, r, i, z, …, to make sure I did not misspell anything. Why were the two strings different? Once again, Sherlock Holmes was right:
When you have eliminated the impossible, whatever remains, however improbable, must be the truth.
I compared every single letter, but omitted the colon.
Lessons learned:
It is unbelievable that this simple mistake took me more than two hours to fix. Part of the reason was that I was a little skeptical of what Jeroen told me:
handle_setheaders()
worked perfectly for me, whereashttpheader
did not—how could they be equivalent? I was misled by the “fact” that I observed, and even spent some time on reading the source code to check if they are really equivalent. I was simply wasting my time. I should never doubt Jeroen again about his own packages! When two things look apparently identical, they may not be. When in doubt, use==
instead of my eyes. -
I clearly knew from the very beginning of the blogdown development that users would hate the fact that they could not click on the Knit button to compile a post. That was due to a technical limitation that I thought was impossible to conquer, after a communication with a colleague. After another communication three years later (because it was such a painful problem), I realized that I misunderstood the limitation and overestimated the difficulty to remove it. By that time, I had suffered a lot from it as the developer, and so did blogdown users.
Lessons learned:
When a pain is foreseeable, try harder to fix it directly (which may involve trying harder to communicate), instead of trying to provide a non-intuitive workaround.
-
I made a change in knitr that should have been made right in the beginning ten years ago. In hindsight, it is obvious that I should require the backticks in chunk delimiters to match, but this requirement had not existed in the ten years. I thought few people would have unbalanced chunk delimiters in their R Markdown documents, but I was wrong and there were all sorts of unbalanced delimiters in the wild. I made two mistakes here: 1) I made the change too quickly, and ended up patching knitr on CRAN immediately on the next day; 2) I did not work hard enough on the error message—originally when the chunk delimiters are not balanced, users will run into an obscure error:
Error in parse(text = x, srcfile = src) : attempt to use zero-length variable name
I bet no one understands this error. Later I made the error message much clearer, telling users exactly which lines in the R Markdown document contain unbalanced chunk delimiters.
Lessons learned:
Be more patient with testing reverse dependencies of a package. The breakage does not happen often, but when it happens and I fail to realize it in advance, I’d be in a great hurry to put out the fire. Work harder to avoid obscure error messages.
To be continued…