Floating point rounding error in _guess_delimiter of csv module

[paulbakkerbloom]

Bug report

Bug description:

Hi! Using the csv module - when determining the delimiter in the _guess_delimiter method, it checks the "consistency" (comparison of counts of a character between rows) of a character. It does reading chunks of 10 lines.

So for the first chunk, when a ';' appears exactly e.g. 30 times on every row - the consistency is 100% - it passes straight away and ';' is recognised as the delimiter. When it appears 30 times on 9 rows and 29 times on 1 row, the consistency is 90%.

The bug is as follows: the required consistency starts at 100% and is lowered by 1% for some iterations until it reaches the minimum threshold of 90%. Only due to floating point rounding errors the consistency becomes 0.8999999, therefore missing the last iteration.

 consistency = 1.0
 # minimum consistency threshold
 threshold = 0.9
 while len(delims) == 0 and consistency >= threshold:
 for k, v in modeList:
 if v[0] > 0 and v[1] > 0:
 if ((v[1]/total) >= consistency and
 (delimiters is None or k in delimiters)):
 delims[k] = v
 consistency -= 0.01

Image with "proof":

Whole method:

 def _guess_delimiter(self, data, delimiters):
 """
 The delimiter /should/ occur the same number of times on
 each row. However, due to malformed data, it may not. We don't want
 an all or nothing approach, so we allow for small variations in this
 number.
 1) build a table of the frequency of each character on every line.
 2) build a table of frequencies of this frequency (meta-frequency?),
 e.g. 'x occurred 5 times in 10 rows, 6 times in 1000 rows,
 7 times in 2 rows'
 3) use the mode of the meta-frequency to determine the /expected/
 frequency for that character
 4) find out how often the character actually meets that goal
 5) the character that best meets its goal is the delimiter
 For performance reasons, the data is evaluated in chunks, so it can
 try and evaluate the smallest portion of the data possible, evaluating
 additional chunks as necessary.
 """

 data = list(filter(None, data.split('\n')))

 ascii = [chr(c) for c in range(127)] # 7-bit ASCII

 # build frequency tables
 chunkLength = min(10, len(data))
 iteration = 0
 charFrequency = {}
 modes = {}
 delims = {}
 start, end = 0, chunkLength
 while start < len(data):
 iteration += 1
 for line in data[start:end]:
 for char in ascii:
 metaFrequency = charFrequency.get(char, {})
 # must count even if frequency is 0
 freq = line.count(char)
 # value is the mode
 metaFrequency[freq] = metaFrequency.get(freq, 0) + 1
 charFrequency[char] = metaFrequency

 for char in charFrequency.keys():
 items = list(charFrequency[char].items())
 if len(items) == 1 and items[0][0] == 0:
 continue
 # get the mode of the frequencies
 if len(items) > 1:
 modes[char] = max(items, key=lambda x: x[1])
 # adjust the mode - subtract the sum of all
 # other frequencies
 items.remove(modes[char])
 modes[char] = (modes[char][0], modes[char][1]
 - sum(item[1] for item in items))
 else:
 modes[char] = items[0]

 # build a list of possible delimiters
 modeList = modes.items()
 total = float(min(chunkLength * iteration, len(data)))
 # (rows of consistent data) / (number of rows) = 100%
 consistency = 1.0
 # minimum consistency threshold
 threshold = 0.9
 while len(delims) == 0 and consistency >= threshold:
 for k, v in modeList:
 if v[0] > 0 and v[1] > 0:
 if ((v[1]/total) >= consistency and
 (delimiters is None or k in delimiters)):
 delims[k] = v
 consistency -= 0.01

 if len(delims) == 1:
 delim = list(delims.keys())[0]
 skipinitialspace = (data[0].count(delim) ==
 data[0].count("%c " % delim))
 return (delim, skipinitialspace)

 # analyze another chunkLength lines
 start = end
 end += chunkLength

 if not delims:
 return ('', 0)

 # if there's more than one, fall back to a 'preferred' list
 if len(delims) > 1:
 for d in self.preferred:
 if d in delims.keys():
 skipinitialspace = (data[0].count(d) ==
 data[0].count("%c " % d))
 return (d, skipinitialspace)

 # nothing else indicates a preference, pick the character that
 # dominates(?)
 items = [(v,k) for (k,v) in delims.items()]
 items.sort()
 delim = items[-1][1]

 skipinitialspace = (data[0].count(delim) ==
 data[0].count("%c " % delim))
 return (delim, skipinitialspace)

CPython versions tested on:

3.10

Operating systems tested on:

macOS

Linked PRs

• [3.15] gh-111487: Fix csv.Sniffer never testing the minimum delimiter consistency #154872

[rhettinger]

Thanks for the bug report. You've done a good analysis of the behavior.

We could use exact fractional arithmetic in this section. However, the threshold of 90% was arbitrary so it likely doesn't make sense to make accumulation more exact. Also, we should be cautious about making any changes to guessing or sniffing logic. Code that is currently working but is on the margins may stop working.

[paulbakkerbloom]

Hi! Well for what it's worth as you can see from the screenshot it also already happens at 93%.

And even though I get that the threshold is arbitrary - now that it's set I do think it has value to stick to it. Simply because than the code acts the way you expect when you read it. It took me some time to figure out why this was going wrong.

If you don't want to change anything to the current logic because it might break existing stuff (not sure if I agree with this), maybe at least we should make this explicit and change to > instead of >=

[serhiy-storchaka]

Fixed in 3.16 by the csv.Sniffer rewrite (gh-83273): _guess_delimiter() and its threshold loop no longer exist, and a sample with 9 rows containing 30 delimiters and one row containing 29 is now detected instead of raising csv.Error.

The maintenance branches are still affected, and a targeted fix for them is welcome.