Title: "fnmatch" exponential execution time
Type: performance Stage: resolved
Components: Library (Lib) Versions: Python 3.9
Status: closed Resolution: fixed
Dependencies: Superseder:
Assigned To: tim.peters Nosy List: Anthony Sottile, kleshni, nedbat, tim.peters
Priority: normal Keywords: patch

Created on 2020-05-03 09:37 by kleshni, last changed 2020-05-12 14:54 by tim.peters. This issue is now closed.

Pull Requests
URL Status Linked Edit
PR 19908 merged tim.peters, 2020-05-04 18:45
PR 20049 merged tim.peters, 2020-05-12 01:13
Messages (13)
msg367963 - (view) Author: (kleshni) Date: 2020-05-03 09:37
Hello. The following code hangs:

import fnmatch
fnmatch.fnmatchcase("a" * 32, "*" * 16 + "b")

Measurements show that its execution time rises exponentially with the number of asterisks. Bash and SQLite 3 process this glob instantly.

This is because "fnmatchcase" generates a regular expression with repeated dots:


It's equivalent to:


But works in exponential time. So the solution is to replace multiple asterisks with a single one, so the latter regexp is generated instead.
msg367982 - (view) Author: (kleshni) Date: 2020-05-03 16:42
>So the solution is to replace multiple asterisks with a single one
No, this is not a solution. This glob also hangs:

fnmatch.fnmatchcase("a" * 32, "*a" * 16 + "b")

And again, Bash and SQLite 3 work perfectly.
msg367983 - (view) Author: Tim Peters (tim.peters) * (Python committer) Date: 2020-05-03 17:06
Note that doctest has the same kind of potential problem with matching ellipsis (0 or more characters) in expected output blocks.  Backtracking isn't needed at all to resolve patterns of that limited kind, but I don't think Python's re module supports enough gimmicks to disable backtracking.

So instead doctest has its own


function to do it in worst-case linear time.
msg368002 - (view) Author: Tim Peters (tim.peters) * (Python committer) Date: 2020-05-04 00:00
"The trick" with this kind of pattern is that a `*` should consume as little as possible to allow the next fixed portion to match.  Apply that at every stage, and it succeeds or it doesn't.  Backtracking can't change that outcome.  For, e.g., the shell pattern:


a regexp to do this without backtracking would be like this, IF Python's re engine supported "atomic groups" (but it doesn't):


The same effect can be gotten in a more convoluted way, via positive lookahead assertions and backreferencing (which Python's re engine does support):


Assertions are "one and done":  if the overall match fails, assertions don't try alternatives.

So that's _a_ way to proceed.  I'm unclear on that it's worth it, though.  Stuff like "*a*a*a*a*a*a*" is just hard to swallow as a shell pattern that would occur in real life.
msg368075 - (view) Author: Tim Peters (tim.peters) * (Python committer) Date: 2020-05-04 18:40
Changed version to 3.9, because anything done would change the regexp generated, and fnmatch.translate()` makes that regexp visible.
msg368222 - (view) Author: Tim Peters (tim.peters) * (Python committer) Date: 2020-05-06 02:28
New changeset b9c46a2c2d7fc68457bff641f78932d66f5e5f59 by Tim Peters in branch 'master':
bpo-40480 "fnmatch" exponential execution time (GH-19908)
msg368645 - (view) Author: Ned Batchelder (nedbat) * (Python triager) Date: 2020-05-11 20:02
This change has caused a problem for  The full details are here: combines fnmatch-produced regexes by joining them with pipes into one larger regex.  The \<g1> group in the regexes now makes that larger regex invalid.
msg368655 - (view) Author: Tim Peters (tim.peters) * (Python committer) Date: 2020-05-11 21:24
Ned, would it be possible to rewrite code of the form:

    if giant pasted regexp matches:

    if any(p matches for p in patterns):

That should work under any version of Python.

There's no guarantee that regexps _can_ be pasted together and still work, so I can't call this change "a bug".  That pasting regexps together "worked" before was an implementation accident.

I'd be happy to change it anyway, except I know of no way to use Python's re engine without backreferences that can avoid exponential-time behavior in some cases.  In some other regexp engines, yes (e.g., as the code comments note, in those that support "atomic grouping"), but not in Python's.  Nor does Python's re engine support reusing backreference names or numbers.

So I know of no way to restore the ability to paste regexps together that wouldn't reintroduce the possiblity of exponential time failure :-(
msg368683 - (view) Author: Anthony Sottile (Anthony Sottile) * Date: 2020-05-11 23:52
one way might be to give the groups "unique" names (perhaps hashing the input string?)  ((this is what I attempted to do in a little bit of code which tried to "backport" (group)*+ and (group)++))
msg368686 - (view) Author: Tim Peters (tim.peters) * (Python committer) Date: 2020-05-12 00:12
I don't want something probabilistic.  Fix it or don't ;-)

One thing that would work, but at the cost of non-determinism:  do the same as now, but obtain the number part of the group name by applying next() to a module-global private instance of itertools.count().  That will keep the numbers increasing "forever", and across calls.  The point to using .count() is that it's atomic (i.e., won't repeat a number if multiple threads happen to be constructing regexps simultaneously).

It's a darned silly amount of effort, though ;-)
msg368695 - (view) Author: Tim Peters (tim.peters) * (Python committer) Date: 2020-05-12 02:19
New changeset b1b4c790e7d3b5f4244450aefe3d8f01710c13f7 by Tim Peters in branch 'master':
bpo-40480: restore ability to join fnmatch.translate() results (GH-20049)
msg368714 - (view) Author: Ned Batchelder (nedbat) * (Python triager) Date: 2020-05-12 09:46
Wow, thanks Tim. To be honest, I was coming around to your original point of view that it was too much to promise that these regexes could be combined the way I'm doing it.  If we have to undo this latest change, I can live with it.
msg368734 - (view) Author: Tim Peters (tim.peters) * (Python committer) Date: 2020-05-12 14:54
Ned, I'm happy to do this. While the ability to join wasn't documented, it's not an unreasonable expectation. I'm not sure it's possible to fail _unless_ the regexps use named groups (and/or numbered backreferences) - and nobody in their right mind would expect regexps for such simple patterns to do such a thing ;-)

So chances seem decent the regression your user stumbled into wouldn't be the only one to pop up.  The fix was actually quite easy (and thanks to Anthony for nudging me in that direction!).  The annoying part was writing a test given that the precise group names generated are no longer predictable :-(
Date User Action Args
2020-05-12 14:54:46tim.peterssetmessages: + msg368734
2020-05-12 09:46:13nedbatsetmessages: + msg368714
2020-05-12 02:19:41tim.peterssetmessages: + msg368695
2020-05-12 01:13:09tim.peterssetpull_requests: + pull_request19358
2020-05-12 00:12:46tim.peterssetmessages: + msg368686
2020-05-11 23:52:18Anthony Sottilesetnosy: + Anthony Sottile
messages: + msg368683
2020-05-11 21:24:20tim.peterssetmessages: + msg368655
2020-05-11 20:02:25nedbatsetnosy: + nedbat
messages: + msg368645
2020-05-06 02:31:54tim.peterssetstatus: open -> closed
assignee: tim.peters
resolution: fixed
stage: patch review -> resolved
2020-05-06 02:28:31tim.peterssetmessages: + msg368222
2020-05-04 18:45:49tim.peterssetkeywords: + patch
stage: needs patch -> patch review
pull_requests: + pull_request19223
2020-05-04 18:40:10tim.peterssetstage: needs patch
messages: + msg368075
versions: + Python 3.9, - Python 3.8
2020-05-04 00:00:38tim.peterssetmessages: + msg368002
2020-05-03 17:06:41tim.peterssetnosy: + tim.peters
messages: + msg367983
2020-05-03 16:42:59kleshnisetmessages: + msg367982
2020-05-03 09:37:07kleshnicreate