在 8 亿行中获取按日期排序的列的唯一性

Question

Input : Multiple csv with the same columns (800 million rows) [Time Stamp, User ID, Col1, Col2, Col3]

Memory available : 60GB of RAM and 24 core CPU

Input Output example

Problem : I want to group by User ID, sort by TimeStamp and take a unique of Col1 but dropping duplicates while retaining the order based on the TimeStamp.

Solutions Tried :

1. Tried using joblib to load csv in parallel and use pandas to sort and write to csv (Get an error at the sorting step)
2. Used dask (New to Dask); \

LocalCluster(dashboard_address=f':{port}', n_workers=4, threads_per_worker=4, memory_limit='7GB') ## Cannot use the full 60 gigs as there are others on the server           
ddf = read_csv("/path/*.csv")                                
ddf = ddf.set_index("Time Stamp")                                        
ddf.to_csv("/outdir/")

Questions :

1. Assuming dask will use disk to sort and write the multipart output, will it preserve the order after I read the output using read_csv ?
2. How do I achieve the 2 part of the problem in dask. In pandas, I'd just apply and gather results in a new dataframe?

def getUnique(user_group):  ## assuming the rows for each user are sorted by timestamp
  res = list()
  for val in user_group["Col1"]:
    if val not in res:
      res.append(val)
  return res

Please direct me if there is a better alternative to dask.

Answer 1

So, I think I would approach this with two passes. In the first pass, I would look to run though all the csv files and build a data structure to hold the keys of user_id and col1 and the "best" timestamp. In this case, "best" will be the lowest.

Note: the use of dictionaries here only serves to clarify what we are attempting to do and if performance or memory was an issue, I would first look to reimplement without them where possible.

so, starting with CSV data like:

[
    {"user_id": 1, "col1": "a", "timestamp": 1},
    {"user_id": 1, "col1": "a", "timestamp": 2},
    {"user_id": 1, "col1": "b", "timestamp": 4},
    {"user_id": 1, "col1": "c", "timestamp": 3},
]

After processing all the csv files I hope to have an interim representation of:

{
    1: {'a': 1, 'b': 4, 'c': 3}
}

Note that a representation like this could be created in parallel for each CSV and then re-distilled into a final interim representation via a pass 1.5 if you wanted to do that.

Now we can create a final representation based on the keys of this nested structure sorted by the inner most value. Giving us:

[
    {'user_id': 1, 'col1': ['a', 'c', 'b']}
]

Here is how I might first approach this task before tweaking things for performance.

import csv

all_csv_files = [
    "some.csv",
    "bunch.csv",
    "of.csv",
    "files.csv",
]

data = {}
for csv_file in all_csv_files:
    #with open(csv_file, "r") as file_in:
    #    rows = csv.DictReader(file_in)

    ## ----------------------------
    ## demo data
    ## ----------------------------
    rows = [
        {"user_id": 1, "col1": "a", "timestamp": 1},
        {"user_id": 1, "col1": "a", "timestamp": 2},
        {"user_id": 1, "col1": "b", "timestamp": 4},
        {"user_id": 1, "col1": "c", "timestamp": 3},
    ]
    ## ----------------------------

    ## ----------------------------
    ## First pass to determine the "best" timestamp
    ## for a user_id/col1
    ## ----------------------------
    for row in rows:
        user_id = row['user_id']
        col1 = row['col1']
        ts_new = row['timestamp']
        ts_old = (
            data
                .setdefault(user_id, {})
                .setdefault(col1, ts_new)
        )

        if ts_new < ts_old:
            data[user_id][col1] = ts_new
    ## ----------------------------

print(data)

## ----------------------------
## second pass to set order of col1 for a given user_id
## ----------------------------
data_out = [
    {
        "user_id": outer_key,
        "col1": [
            inner_kvp[0]
            for inner_kvp
            in sorted(outer_value.items(), key=lambda v: v[1])
        ]
    }
    for outer_key, outer_value
    in data.items() 
]
## ----------------------------

print(data_out)