Preface (revised early 2015). Some of the following article is subjective and necessarily personal, requiring a first-person tone. It has also grown a bit dated, but I've retained it here in hope that it may still be able to stimulate some much-needed discussion.
If anything, some of the questions raised here have only grown more acute since I wrote this. For example:
For more recent thoughts on Python's own evolution—the main theme of this article—see the introduction and conclusion to 2013's Learning Python, 5th Edition (Chapters 1 and 41); the continuing saga of feature growth and change in Python 3.4; a short 2013 blog piece I did for O'Reilly on new-style inheritance available here and here; and a related 2015 post on Python 3.5 type hints.
January, 2012 (last substantial edit September, 2012)
It's official: I've now been using and advocating the Python computer programming language for two decades. I have no idea where the time went (the current theory is that much of it wound up on the West Coast). It's been a busy time. All told, I've spent 15 of those years teaching 250 live Python classes, and 17 years writing 4 editions each of my 3 Python books. Despite the passage of time, I look forward to teaching and writing about Python for years to come. After two decades, writing Python scripts is still too much fun to stop.
That said, it's impossible to not be at least a little reflective when reaching a major milestone in any career path. I've watched Python grow from the new kid on the scripting languages block in 1992 to one of the top programming languages used in the work today. Especially given Python's much larger user baser and its presently divisive 2.X/3.X versions fork, it's only natural to wonder what's next. I might mull over such things more than some, of course, but my long sojourn as Python programmer, teacher, and advocate probably affords me a somewhat unique perspective on them.
With the benefit of some hard-won Python historical context, then, here are a few observations about Python's future. To me, three questions seem paramount to it, and capture both the dilemmas and opportunities facing Python today:
Has Python grown too complex, or is it only as complex as it must be given its many roles, and still much simpler than alternatives like Java, C#, and C++?
A simple line and character count would convince most observers that Python is still much easier to use than most of its contemporaries. Thanks to its dynamic typing, built-in toolset, and minimal syntax, most Python scripts continue to weigh in at 1/3 to 1/5 as much code as they'd be in Java or C++. You can get much more done in Python with much less typing, and you can read your code after you do so.
This is as true a distinction for Python today as it was at its beginning, and is undoubtedly one of the main reasons that Python continues to attract such a wide audience. Its relative ease of use remains a compelling feature to many—from curious newcomers and occasional scripters, to avid hobbyists and professional developers.
Still, with the addition of tools like generators, context managers, decorators, and Unicode, Python has clearly grown more complex over the years. One could even make a case that it may have lost some of its original focus on simple scripting along the way, a model now championed by newer arrivals on the programming language scene such as Lua.
The upside to Python's growth is that it is substantially more powerful today, and better suited to larger-scale systems development than simpler tools. It seems to fill a midlevel niche between scripting languages (like Lua and Tcl) and systems languages (like Java and C++), but can still usually serve in both roles if needed. With support for procedural, object-oriented, and function programming paradigms, Python's present scope is broader than it was in the past.
But the problem with adding powerful new tools is that you must add powerful new tools—and burden your new and existing users with them in the process. This has both intellectual and pragmatic consequences. As I point out in a sidebar at the end of the latest Learning Python (and in the conclusion chapter of its newer 5th Edition), as soon as one person uses such a new or advanced tool in their code, it's no longer optional for anyone who touches that code. It raises the intellectual bar for most of the community.
And while Python is very broadly used, some systems which may have deployed it as a scripting layer in the past are today choosing lighter-weight options. See the gaming domain for notable examples; whether warranted or not, there is a growing conception that Python has become too complex or large for some potential embedded application roles that it once may have addressed. More features mean more power, but also less scope.2
I've heard both sides of this issue from Python newcomers and veterans alike. While some applaud new language features, others may be hesitant to base a project on what they perceive as shifting sands, or worse, bloat-ware. Most agree that recent growth has nudged Python's scope towards larger scale development and away from simple scripting tasks, though they may differ on the degree and merit of that shift.
Whether Python is viewed as a simpler Java, a more powerful Lua, or a class unto itself, its focus and identity are bound to affect its reach in years to come. What Python is has always been up to its users, of course, but we'd do well to safeguard the identity and scope that made Python successful in the first place. Feature creep is also related to the broader issue of stability, and next question.
Should Python developers strive for a stable base that its current users can rely on, or pursue a more dynamic evolution independent of those users?
Python's long history of backward compatibility came to a screeching halt in late 2008 with the release of version 3.0, an intentionally incompatible version of the language intended to resolve longstanding legacy concerns. Even within the 2.X and 3.X lines, though, change has always been constant; some of it broke existing code, and not all of it has been in the name of fixing genuine problems. This seems almost indigenous to open source projects, but it has grown a large enough concern among Python newcomers to merit careful consideration by Python's developers.
By most ballpark estimates, there are roughly 1M Python users (plus or minus a few). By contrast, some 1K people attend PyCon conferences, and the number of active Python core developers is on the order of just 100. That means the user base outnumbers conference goers by 1,000 to 1, and developers by 10,000 to 1. Put another way, conferences and core developers reflect just 0.10% and 0.01% of the Python user base, respectively.3
This is as it must be: not everyone can attend conferences, and Python maintenance is a volunteer effort. But when these smaller groups make changes that impact all, this seems a case of the few dictating to the many in the extreme. And when these few break existing code, there must be very clear cause if they hope to keep the many in their camp.
This massive difference in scale between Python users and Python developers is also a likely large part of the reason that the audience for the non-backward compatible Python 3.X is still just a fraction of the 2.X audience, more than 3 years after 3.X's release, and more than half way through the 5 year worst case projection for 3.X's rise to dominance. The user base seems to be voting with its downloads and ports (or lack thereof).
As one metric, Python 2.X is still downloaded for Windows at python.org roughly but consistently 3 times as often as Python 3.X as of today, January 2012. By some analyses, the actual 2.X/3.X gap is likely to be even wider than this 3:1 ratio implies.4 But even with a positive spin, 3.X now has less than 2 years left on its 5 year plan, to overcome the current 3:1 spread and subsume 2.X. That may be asking a lot.
As another measure, many popular systems still aren't yet available in 3.X form, including Django and wxPython; see python.org's survey for an up-to-date list. Of note, Google's popular App Engine Web development system is also 2.X-only today, with an upcoming upgrade based on Python 2.7. Python 3.X's potential to supersede 2.X is still unclear, well into the second half of this milestone's projected timeframe.
Ironically, the most plausible explanation for this slow (or limited) adoption is Python 2.X's success. It may be that 2.X is too widely used and 3.X is too much of a leap for the shift to occur. To some, it seems increasingly likely that 2.X and 3.X will coexist indefinitely, as distinct projects—both active and widely used, but receiving fewer developer cycles than a single solution might. You can't often pull a well-worn rug out from under 1M users; they won't gladly leap to another, and are usually too heavy to move in any event.
On the other hand, some view change as an asset that keeps Python up to date and relevant. Adding new features to a tool can help avoid decline in a marketplace of ideas as dynamic as the software field (this is also why Python standardization efforts have never gotten off the ground5). Moreover, some have long seen Python as both a sandbox for exploring new ideas, and a sort of teaching project for newcomers to cut their large-scale software development teeth on. In fact, many in the Python developers group have explicitly encouraged relative beginners to get involved for this very purpose.
But is this really the best context for amateurs to learn from their mistakes? There seems an inherent danger in allowing the less experienced to experiment by changing a tool which so many rely on: focus and quality can both sometimes suffer. See this post for one recent quality example in Python 3.2—a temporary breakage of Python's input() function on Windows, a fundamental and widely used tool. The initial I/O performance problems in Python 3.0 described here may also fall into this category—though made a moot point by later 3.X releases, 3.0 was so weak on this front (up to 1,000 times slower than 2.X in worst cases) as to be unusable in some contexts, and considered by many to be beta quality code.
All software has bugs, of course (including mine), and most software evolves over time (including mine). But there is a social responsibility issue here too: software that is as widely used as Python merits extra caution, lest quality issues alienate its user base. In the case of the 3.2 input() breakage, many of the posts I've seen suggest dumping 3.X and going back to 2.X—probably not the outcome that 3.X's developers intended.
Bugs aside, some of Python's recent changes also underscore how its focus can seem to shift according to the whims of its current developers. Its new emphasis on functional programming tools like generators and closures, for example, makes Python more powerful, but also more difficult to learn for those without backgrounds in this model. Python becomes a multi-paradigm tool, with multi-paradigm prerequisites—its users today must generally master both object-oriented and functional techniques, if they are to understand all the code they are likely to encounter in the wild. And its users today must accept that new paradigms might creep into the language tomorrow, and base projects on what might be a perpetually moving target.
Even much narrower changes made by Python's developers can sometimes seem to stem from little more than the personal preferences of the initiator. See this post for some recent examples in this department—a change to the struct module which removed existing and documented functionality, and a cgi module change that relocated a very widely used tool—both of which occurred long after 3.0's open season of incompatible change, and serve as recent representatives of a much broader category of Python change.
I don't mean to unfairly single out these two changes or their advocates (some of whom I count as personal friends) and could list dozens of similar examples, but these two reflect a culture that seems to surround Python development today. In both cases, the modifications made will break existing code, not in the name of fixing bugs, but only for the sake of the aesthetic whims of individuals or very small groups working with virtually no input from the user base at large.
One could also argue the merit of these changes, of course: the automatic UTF-8 encoding in struct may seem inconsistent to some (though in 3.0 through 3.2, DBM files also auto-encode str keys per UTF-8, and ftplib auto-encodes lines per Latin-1), and HTML escaping has roles beyond CGI scripting. But it shouldn't matter—in a tool as widely used as Python, once something is supported in released code, and is formally documented as being supported in the manuals, the window of retrospective analysis should be closed except in the case of true bugs. How else can anyone rely on such a tool?
Unfortunately, this type of change has been a relative constant in Python's history. Having been on the front lines of Python documentation and teaching since 1995, I've seen the negative impacts of rapid change first-hand. Especially in the larger Python world of today, what may seem interesting to core developers often comes across as arbitrary and even aggravating to the user base.
These people aren't "haters"—a label tossed out sarcastically at a recent Python conference. They are having an honest and reasonable reaction to a crucial issue in this domain: Developers of commercial compilers, and established software in general, do most of their work by responding to requests from actual users, not by initiating requests on their own. Open source projects often seem to follow the opposite path, even when the changes initiated by their developers are incompatible with masses of existing user code.
From the outside, it would be impossible for some not to see this development model as a sort of anarchy at best, and a tyranny of the minority at worst. Such states can flourish in an open source project only under a silent user base, but most people are simply too busy using Python to monitor its developers' actions. The larger consequence of subjective change by the few is to further cloud Python's perceived stability. Barring a standardized version of Python (which, as mentioned earlier, has yet to gain traction), developer restraint seems the only solution here.6
A final argument raised in this department is that users of an open source tool are in control of their own destiny: they can stick with the version they've fetched, and need not upgrade to new, incompatible releases. This seems a bit thin to me. If users never upgrade as a tool evolves, they soon become badly out of date with libraries, tools, on-line help, and books. This is true even if users maintain their Python themselves: while their dated code might run in some newer Pythons, they will still wind up well behind the curve in terms of both resources, and the language's ecosystem at large.
Python's history is a good case in point here: code written in Python today looks very different than it did in the past. In fact, I commonly write scripts which contain almost no lines that would have run in a Python of just a few years ago. If users had installed 2.0, for instance, and never upgraded, their scripts would be badly out of date today, and nearly every available learning resource would be largely irrelevant to their code. As is, today's users must take care with books to select from one that covers 2.X, 3.X, or both. Even in open source, users do have to keep up with new releases, just as they must keep up with new language features, or risk being alienated from most forms of support.
This stability-versus-relevance debate reflects a basic difference of philosophies, and is a core dilemma of open source in general. As a former compiler developer I know that experimenting with language features can be fun. But as an author and trainer who has taught thousands of Python users in person, I also know that stability becomes crucial once a system grows to be as widely used as Python. For better or worse, this tradeoff bears directly on the choice of tools like Python in production environments, and will likely continue to be a factor in Python's future. It also impacts the size of the learning task for Python newcomers, the next and final question's theme.
Do some think that Python is easier to learn than it is, or have they simply been misled by the unrealistic claims of marketing?
As a trainer and writer, I've confronted this question directly and personally for some time. It has only grown more acute as Python has shed some of its original simplicity over the years. I took a first pass over this topic in a post on my book site, How long will it take to learn Python?, and addressed it in my training site's Mission statement, so I won't repeat those documents' learning curve notes in full here.
In short, it seems time to start setting more realistic expectations about the effort required to learn a programming language like Python. This is true both in the publishing and training industries, and in the software field at large. As I'll discuss here, understating the entry requirements and learning curve in such a profession might temporarily benefit salary payers and peddlers of inadaquate education offerings; but it won't work in the long run.
Let's be honest: programming is both highly rewarding and deeply challenging. Except for very limited scripting contexts, software development takes years to master. Downplaying—or deliberately misstating—this does a disservice both to newcomers and to the field at large. On the one hand, a misunderstanding of the learning curve can be an obstacle to Python adoption when expectations lead to frustration among beginners. On the other, the software field desperately needs skilled engineers who were willing to invest the time and focus required to master its content; in their absence, quality can become problematic for all stakeholders.
The downside to such brutal realism is that it may scare away some potential or budding programmers who might have otherwise gone on to become talented practitioners. Teaching is as much about inspiring as it is about knowledge transfer. Encouraging newcomers is never wrong, as long as it's honest. After all, we all were beginners once, and many of us owe our careers to such positive reinforcement. We need only be more accurate in our promises.
Some have suggested that a limited subset of Python can still be useful for limited roles—serving roughly the same purpose that BASIC did at the start of the PC era, and inspiring a new generation of novices, many of whom may move on to full proficiency in time. As a prime example, witness Python's role as the official programming language of the Raspberry Pi single-board computer, a device aimed in part at educating nonprogrammers and children. This seems a clearly admirable initiative; excited newcomers are ultimately the future of our field.
But while limited subsets of Python can be had relatively quickly by some, no newcomer can fully master it in 24 hours, 3 days, or any other claim which may sell products but masks the true size of the learning task. Object-oriented programming alone, for example, is beyond the reach of most non-developers in such short timeframes. Moreover, subsets of Python large enough to be truly useful are much more accessible to already-experienced developers than beginners. This may not be welcome news to a generation accustomed to shortcuts in other domains, but we need to be clear that there is no "Software Hero" to be had. Python can be great fun, but programming it well requires substantial effort. It never was an N-hour skim, and its feature set growth in recent years has only amplified this truism.7
Nevertheless, in recent years there has been no shortage of publishers competing to dumb down their books (to the point that some are more entertainment than education), and training providers promising immediate career benefits for their services (while pushing "classes" that largely consist of watching canned videos or cutting and pasting code). In fact, this is almost common practice at some sites selling certificate programs—a resume benchmark which many employers filling worthwhile positions would either ignore, or consider a strike against a candidate (and a possible caution on the ethics front). While jobs might be rewarding and plentiful for experienced Python programmers in the long run, implying that this is somehow automatic before the first lesson is just plain wrong.
More fundamentally, understating software engineering's entry requirements this way can serve to denigrate the entire field. It might also have the effect of pushing down salaries of its practitioners—whether intended or not. Why would anyone promise shortcuts to proficiency in a highly technical field? And why would anyone expect to find them? Like other engineering domains, software is a lot more interesting and substantial than some of the marketing out there implies.
It's important to note that this idea that nonprogrammers should be able to program did not originate with Python, or even scripting languages in general. In fact, it's been around since the dawn of computer science. One of the main design goals of COBOL, a programming language invented in 1959 and still used widely in business applications today, was to provide an English-like syntax so that, to quote one source: "non-programmers—managers, supervisors, and users—could read and understand the code." Noble goal to be sure, but this field has a very long history of confusing the skill level needed to understand individual lines of code with that required to perform full-scale systems development of larger programs. As the next section argues, both roles have merit, but we blur the line between them at the peril of our field.
A counter argument I've heard on this front is that some tasks require less skill, and producing more developers is better than fewer even if they aren't masters of the domain, because this means more applications will be created. Per this argument, not everyone needs to be a proficient programmer to do programming. There is a big difference between software hobbyist and professional—a distinction too often overlooked by the technical press. Crossing the threshold from the former to the latter requires extra knowledge, experience, and sacrifice which some tasks may not require.
To be sure, many software domains owe either their existence or their recent growth to this viewpoint. For example, in some popular Python application domains such as scientific programming, software development is often considered to be a secondary skill, and given limited focus.8 Web programming and system administration are sometimes similarly seen as less technically demanding than other tasks, and ideal scripting domains. If pressed, many Python users today would happily consider themselves more hobbyist than professional, and disinclined to wade any further into the software pond than their specific, narrow goals require.
Such a liberal attitude towards entry requirements may make perfect sense for some code of limited scope. But most programs don't fall into this category; they are used, reused, and then reused again, and often by other than their original authors. More fundamentally, the real danger in blurring the line between hobbyist and professional is that it discounts the huge importance of quality in software: more applications is not an absolute good, if those applications don't work as they should. Poorly written software can become a major liability, both for its users, and for other programmers who must later maintain it. In truth, a widespread lack of quality in applications could prove a much larger negative for the software field than a dearth of applications, and might even contribute to a backlash against computerization in general.
And if you aren't buying that, come back and read this again the next time your game console freezes up; your music player does the wrong thing with your files; your social networking site does the wrong thing with your identity; your car's electronic control system malfunctions; your Blu-ray player chokes on an incompatible disc or device; or your credit card balance is botched by a Web page or an "app"—most of which this author counts as recent personal experience. Quantity doesn't trump quality in this domain. If the software that people depend on doesn't work, people will eventually stop depending on it.9
There are many possible futures, of course, but in one there is a world in which some applications grow so unusable that they are relegated to the cultural fad pile (to be greeted warmly, perhaps, by 8-track players and CB radios). That world may seem unlikely today, but it's dire enough to merit caution when setting expectations and agendas for the next wave of newcomers eager to get into the field.
So there you have it: a few points to ponder from someone who has spent two decades in the Python trenches. I suspect that Python's future may hinge on such questions, but they are too rich to explore further here, where they'll have to remain asked but unanswered. Ultimately they are up to the Python community to resolve over time. I look forward to watching the resolutions continue to unfold.
For my part, I don't believe the sky is falling, do not think Python 3.X will kill Python, and expect that people will continue to be attracted to what still seems a much nicer way to program machines than any of the alternatives. One of the great rewards of teaching Python is the common reaction of students when they realize how much easier their work can be with this tool. While Python might not be the radical paradigm shift that may someday free us from the shackles of the Von Neumann computer architecture altogether, it does represent an incremental step forward, and is still good news for software developers today.
But judging from what I've seen recently of the real world in which people evaluate and learn Python, the answer that seems most perilous is status quo. Mature projects merit mature perspectives. The model normally followed by Python developers in the past—hacking as they wish and letting the chips fall where they may—might be fun and may have worked well in Python's early growth years, but there's a lot more to think about today. Whether Python is a tool for hobbyists or professionals, a project focused on sandbox or stability, or everything to everyone at once, I suspect that its perceived identity will probably be more important to its future than any latest and greatest language feature.
A more thorough retrospective at the two-decade milestone of my Python career would undoubtedly include attack ships on fire off the shoulder of Orion (and that sort of thing), but is thankfully beyond this article's scope.
1 Yes, this article's title is a Monty Python reference.
2 But game programming is done regularly in Python too. A major counterexample in this domain, the popular Eve Online massively multiplayer online game is heavily based on Python, both on the client and server. This seems to use Python as a core development language, though, not in an embedded scripting role. One note about Eve Online's Python use, which also describes its upgrade to 2.7 as well as its reluctance to upgrade to 3.X, lives here. There's also a quote from Eve Online on this page, and much more about game programming in Python here.
3 Interestingly, the ratio of PyCon attendees to Python users at large is roughly the same as the ratio of readers who post book reviews about my books on web sites to the books' customer base at large—some 1 in 1000. The few do seem to drown out the many on the Web.
Update: as of mid 2012, PyCon attendance shot up to 2,500 and sales of my books reached 400K units, but the user-base ratios in this document are still fairly accurate, at least in their order of magnitude (power of 10).
4 I analyze Python downloads at python.org using the site's webstats page, collecting KBs downloaded for the top 2.X and 3.X Windows install files, and dividing by file size to get units. This reflects Windows users only, and is probably skewed to artificially inflate 3.X's popularity, because it doesn't fully account for the large existing 2.X user base, or users of systems with preinstalled 2.X Pythons. Linux, Mac OSX, and many Python-scripted packages typically come with 2.X as their default Python today, and most Python newcomers would probably be more inclined to fetch the latest and greatest version, 3.X. Still, the 3:1 ratio for 2.X:3.X Windows downloads has been fairly consistent for some time, suggesting at least that the adoption rate for 3.X is still slow, regardless of its true magnitude, and nowhere near the 3.X dominance forecast by some.
Update: this 3:1 ratio for 2.X to 3.X downloads at python.org still holds true as of July, 2012.
Update: Python developers stated that 3.X downloads achieved parity with 2.X downloads sometime around late 2013 to early 2014 (but this is unverifiable, as webstats was removed in a python.org redesign at roughly the same time as this claim). Important if true, but hardly earthshattering, some 5 years after 3.X's release!
Update: 2015 still hosts a dual 2.X/3.X world, and radical change in the future now seems unlikely.
5 Though a standardized version of Python might address many of the issues raised in this article: it would provide a stable base which users could rely upon, and could ease the learning process by removing the specter of perpetual change from newcomer curriculums. It's not impossible that Python 2.7, the last of the 2.X line, might become a de facto standard of sorts, due to both 2.X pervasiveness, and the fact that 2.7 is no longer being actively changed. In this scenario, 2.7 would be base Python, and 3.X would be an experimental branch—which is one way to interpret the current 2.X/3.X adoption rates. The future remains to be seen, but 2.X's still much larger user base probably controls it at least as much as 3.X developers.
6 For two more examples in this category, see the just-announced 3.3 decision to deprecate distutils and FTP.nlst(), widely used tools both. You should judge the merits of 3.3's less incompatible yield from... and raise...from None language extensions for yourself, but they may be pushing Python complexity further than some wish. There is good work being done in 3.X too, of course, but some of it will be very difficult for impacted users not to view as arbitrary preference of core developers. I encourage anyone impacted by backward-incompatible changes to register a complaint in Python's development channels. It's your language, after all. For details on reporting such things, read this wiki page. The reporting process seems a bit more difficult than it might be, but is probably worth the effort when changes impact your code. Open source development need only seem like anarchy or tyranny if its users silently accept such a fate.
Update: For another recent example of developer-initiated change made even in spite of end-user practice, see this PyMailGUI patch required for Python 3.3 in 2013; it's a pervasive and ongoing issue.
7 Full disclosure: like most people who wrote books or did training in the software development field in the last 2 decades, I've probably made a few learning curve underestimates myself, though in my defense, mine were intended to be about the curve for already experienced programmers, not true beginners. In fact, the latter group was explicitly removed from the target audience of my books' content early on once it became clear how distinct their needs were (and very shortly after I started receiving emails asking for help running a text editor). Still, misconceptions can be difficult to avoid, especially those that have permeated a field since its inception. For another take on this subject, see also my earlier 2009 post on misleading publisher marketing: Focus, "2.0".
8 Numeric and scientific programming is a major domain for Python; for a quick sample, see this page. For an example of the low esteem in which many in the scientific community hold programming, see the first bullet on this page: "Scientists first, developers second." I don't mean to single out this source unfairly either—Enthought provides a popular turnkey Python distribution which includes numeric and scientific programming extensions, and offers training specific to this domain which I often recommend as a follow-up to my general Python class—but this also seems representative. A student in a Python class I once taught at Los Alamos came up during a break to tell me that "Scientists can program, but programmers cannot do science." When I saw his code later, I realized he was probably wrong on at least one count.
9 It's easy to find evidence on the Web of just how far the underestimation of requirements for careers in software engineering seems to have spread, and how broad the challenge to quality may be. Here's a typical example from a blog picked at random (and with no offense intended to its participants); the word on the Internet street seems to be that amateurs can get into programming without any formal training, in part by contributing to an open source project like Python to flesh out their resumes. As for the blog's other advice, I'll let you ponder the merit of laptop stickers and coffee shops as career advancement tools. We all started out as hobbyists or amateurs, of course. Most of us also understood the importance of education and experience in a technical field.
Latest substantial revision: September 25, 2012 (first posted: January 12, 2012).
Have a comment on this article? send an email.
|Mark Lutz||Books site||Training site|