Contributing a CSV module [RE: dask dataframe `read_csv`]
See original GitHub issueFollowing our discussion in dask #8045 I’m seeking to contributing a CSV module to fsspec-reference-maker
, and wanted to kick off a feature proposal issue here to clarify some aspects.
Hope this is a decent start, but WIP and comments appreciated (I need to review the role of fsspec
in dask more closely).
Firstly, as described in #7, this library is geared towards evenly spaced chunks. The README of this repo gives an example of this with a single variable, or “dimension”, i
(the number of chunks), which is used in the example spec along with a fixed length (1000) to produce range(0, i * 1000, 1000)
.
This of course matches how dask reads CSVs as range(0, size, blocksize)
, but the whole point of my intervention here is to make them no longer evenly spaced deterministic chunks.
The word ‘deterministic’ here seems to mean both:
- “based on the byte positions rather than byte content at those positions”, as well as
- “uniquely identifiable, not changing upon recalculation”.
The alternative I am proposing for CSVs to dask only fits the 2nd of these criteria (as it will involve checking bytes, which may result in changes to the offsets, thus chunks may not stay evenly spaced).
Also in #7 there is mention of “callbacks”, and I’m uncertain whether this is the route I should take to achieving this adjustment to the offsets (or something entirely different I should ignore).
I am inclined to copy the conclusions of that issue, that perhaps it is easiest to begin by aiming to produce the explicit ‘version 0’ spec rather than the templated ‘version 1’ until how this works is clearer to me.
As a less significant problem for me, the docs here refer to jinja2
rendered strings, but I don’t see that library as a requirement anywhere here, so I’m not sure how that works (perhaps it is a future feature, I’m noting this library is a recent/future-facing effort).
Here’s my first attempt at an idea of how this would look (filenames vaguely copying the format of the WIT dataset as “base_0_ordinal_count
of base_1_cardinal_total
”):
-
The values in the 2nd and 3rd parts of these values are supposed to indicate where a previous routine has calculated the offsets (as
10, 5, 20, 0, 50, 25
), which are added to the evenly spaced offsets (1000+10, 2000+5 etc.) and subtracted from the lengths between consecutive offsets (1010-0 = 1010, 2005-1010=995, etc.) -
I’m fairly sure that the 3rd item in the gen keys should indicate the length of the chunk that ends at that offset but please correct me if I’m wrong.
This then gives the spec of a ‘filesystem for partitions within a file’, addressable by filepath plus ‘virtual’ partition index:
{
"key0": "data",
"gen_key0": ["/path/to/csv_0_of_1.csv.gz/partition_0", 1010, 1010],
"gen_key1": ["/path/to/csv_0_of_1.csv.gz/partition_1", 2005, 995],
"gen_key2": ["/path/to/csv_0_of_1.csv.gz/partition_2", 3020, 1015],
"gen_key3": ["/path/to/csv_0_of_1.csv.gz/partition_3", 4000, 980],
"gen_key4": ["/path/to/csv_0_of_1.csv.gz/partition_4", 5050, 1050],
"gen_key5": ["/path/to/csv_0_of_1.csv.gz/partition_5", 6025, 975]
}
- I’m not sure what to put in the ‘key’ entries so have removed the ones from the example spec (please let me know if this is unadvisable, and if you have an idea of what should go there instead)
- I presume the one that is currently bearing the bytes
b”data”
should be storing something important to identify the CSV, but I can’t determine what that is on my first attempt- My understanding is that this will be fed into the
OpenFile
object as thefs
argument, so it should store things relevant to that. Perhapspath
? I’m very unsure how this should look though, and suspect if I guess I’d only end up putting irrelevant info in that’ll already be passed in.
- My understanding is that this will be fed into the
- For simplicity I’m considering 2 files here, each with 3 offsets (i.e. 4 chunks: the offset starting at 0 is always going to be assumed to be valid: if it’s not then that’s a corrupt CSV, not the problem I’m seeking to solve here)
As for the matter of identifying the offset adjustments (10, 5, 20, 0, 50, 25
) I expect the fastest way to do so is
- initialise
separator_skip_count = separator_skip_offset = 0
at each offset mark (1000, 2000, etc.) - try
pandas.read_csv(nrows=1)
- catch failure; increment
separator_skip_count += 1
if it fails (repeat)
- catch failure; increment
- finally [upon success]
- break out of the loop
- use the
tell
minus the offset to give the ‘offset adjustment’ (assignseparator_skip_offset
)- left as 0 for no adjustment (if
separator_skip_count == 0
), or a positive integer
- left as 0 for no adjustment (if
The separator_skip_count
indicating the number of newlines that were skipped after the offset+length to find the ‘genuine’ row-delimiting offset seems redundant to store, but useful while writing/debugging this algorithm.
- I say that, but I don’t know: perhaps it’d be inexpensive to recalculate the actual byte offsets from the number of newlines to skip after the offset, rather than store that offset? (Not clear to me yet)
Only the separator_skip_offset
needs to be stored: summed with the offset, in the 2nd item of the values (1010, 2005, etc.)
I think at the point that the separator_skip_offset
is calculated, the ‘version 1’ spec could be computed, to reduce to the above ‘version 0’ spec, as something like:
{
"version": 1,
"templates": {
"u": "/path/to/csv_0_of_1.csv.gz",
"f": "partition_{{c}}"
},
"gen": [
{
"key": "gen_key{{i}}",
"url": "{{u}}/{{f(c=i)}}",
"offset": "{{(i + 1) * 1000}}",
"length": "1000",
"dimensions":
{
"i": {"stop": 5}
}
}
],
"refs": {
"key0": "data",
}
}
- I may be misunderstanding something by putting the filename within the template rather than as a variable
- should
gen_key0
bepartition0
(etc) ? or should this stay asgen_key0
to make clear that it’s generated? - I can’t figure out where the array specifying the
separator_skip_offset
should go (if I put it in thegen.dimensions
key it’ll become a Cartesian product, whereas I want to ‘zip’ it against thei
range…) - Should I change the
gen.url
key from “url” to something else, since it’s expected to refer to a file path not a web resource?
Without understanding how to incorporate the offset adjustments into this template, I don’t think I can write the ‘version 1’ spec at this time, but I hope we might be able to figure it out here.
Issue Analytics
- State:
- Created 2 years ago
- Comments:13 (13 by maintainers)
Top GitHub Comments
Note that you probably want to take dask out of the equation too - it might be where you want the files to be processes eventually, but I think you should be able to find valid offsets without it, simplifying the process (at the expense of no parallelism).
I would not attempt to solve the compression and parsing issues in one go, it would be better to use an uncompressed target at first, I think.