X-Git-Url: https://code.communitydata.science/mediawiki_dump_tools.git/blobdiff_plain/d8d20f670b98507a3dea27c38e5463c1b0b2e6cd..ae870fed0bdca49df3c475220667cb94fa4937c7:/wikiq?ds=sidebyside diff --git a/wikiq b/wikiq index c2a2f12..d966b50 100755 --- a/wikiq +++ b/wikiq @@ -8,6 +8,7 @@ import argparse import sys import os, os.path import re +from datetime import datetime from subprocess import Popen, PIPE from collections import deque @@ -24,13 +25,11 @@ PERSISTENCE_RADIUS=7 from deltas import SequenceMatcher from deltas import SegmentMatcher -from dataclasses import dataclass -import pandas as pd +import dataclasses as dc +from dataclasses import dataclass, make_dataclass import pyarrow as pa import pyarrow.parquet as pq -from typing import List - class PersistMethod: none = 0 sequence = 1 @@ -41,7 +40,6 @@ def calculate_persistence(tokens_added): return(sum([(len(x.revisions)-1) for x in tokens_added]), len(tokens_added)) - class WikiqIterator(): def __init__(self, fh, collapse_user=False): self.fh = fh @@ -189,33 +187,118 @@ class RegexPair(object): temp_dict[self.label] = None # update rev_data with our new columns - rev_data.update(temp_dict) + for k, v in temp_dict.items(): + setattr(rev_data, k, v) + return rev_data -@dataclass -class RevData(): - revid: int +@dataclass() +class RevDataBase(): + revid: int date_time: datetime articleid: int editorid: int title: str namespace: int deleted: bool - text_chars: int - revert: bool - reverteds: list[bool] - sha1: str - text_chars: int - revert: bool - reverteds: list[int] - minor: bool - editor: str - anon: bool - collapsed_revs:int - token_revs:int - tokens_added:int - tokens_removed:int - tokens_window:int + text_chars: int = None + revert: bool = None + reverteds: list[int] = None + sha1: str = None + minor: bool = None + editor: str = None + anon: bool = None + + pa_schema_fields = [ + pa.field("revid", pa.int64()), + pa.field("date_time",pa.timestamp('ms')), + pa.field("articleid",pa.int64()), + pa.field("editorid",pa.int64()), + pa.field("title",pa.string()), + pa.field("namespace",pa.int32()), + pa.field("deleted",pa.bool_()), + pa.field("test_chars",pa.int32()), + pa.field("revert",pa.bool_()), + pa.field("reverteds",pa.list_(pa.int64())), + pa.field("sha1",pa.string()), + pa.field("minor",pa.bool_()), + pa.field("editor",pa.string()), + pa.field("anon",pa.bool_()) + ] + + def to_pyarrow(self): + return dc.astuple(self) + + def to_tsv_row(self): + + row = [] + for f in self.fields(): + val = getattr(self, f.name) + if getattr(self, f.name) is None: + row.append("") + elif f.type == bool: + row.append("TRUE" if val else "FALSE") + + elif f.type == datetime: + row.append(val.strftime('%Y-%m-%d %H:%M:%S')) + + elif f.name in {'editor','title'}: + s = '"' + val + '"' + if f.name in TO_ENCODE: + row.append(quote(str(val))) + + elif f.type == list[int]: + row.append('"' + ",".join([str(x) for x in val]) + '"') + + elif f.type == str: + if f.name in TO_ENCODE: + row.append(quote(str(val))) + else: + row.append(val) + + return '\t'.join(row) + + # def __init__(revid: int, + # date_time: datetime, + # articleid: int, + # editorid: int, + # title: str, + # namespace: int, + # deleted: bool, + # test_chars: int, + # revert: bool, + # reverteds: list[bool], + # sha1: str, + # minor: bool, + # editor: str, + # anon: bool): + + + +@dataclass() +class RevDataCollapse(RevDataBase): + collapsed_revs:int = None + pa_collapsed_revs_schema = pa.field('collapsed_revs',pa.int64()) + pa_schema_fields = RevDataBase.pa_schema_fields + [pa_collapsed_revs_schema] + +@dataclass() +class RevDataPersistence(RevDataBase): + token_revs:int = None + tokens_added:int = None + tokens_removed:int = None + tokens_window:int = None + + pa_persistence_schema_fields = [ + pa.field("token_revs", pa.int64()), + pa.field("tokens_added", pa.int64()), + pa.field("tokens_removed", pa.int64()), + pa.field("tokens_window", pa.int64())] + + pa_schema_fields = RevDataBase.pa_schema_fields + pa_persistence_schema_fields + +@dataclass() +class RevDataCollapsePersistence(RevDataCollapse, RevDataPersistence): + pa_schema_fields = RevDataCollapse.pa_schema_fields + RevDataPersistence.pa_persistence_schema_fields class WikiqParser(): def __init__(self, input_file, output_file, regex_match_revision, regex_match_comment, regex_revision_label, regex_comment_label, collapse_user=False, persist=None, urlencode=False, namespaces = None, revert_radius=15, output_parquet=True, parquet_buffer_size=2000): @@ -227,14 +310,10 @@ class WikiqParser(): self.collapse_user = collapse_user self.persist = persist - self.printed_header = False self.namespaces = [] self.urlencode = urlencode self.revert_radius = revert_radius - self.parquet_buffer = [] - self.parquet_buffer_size = parquet_buffer_size - if namespaces is not None: self.namespace_filter = set(namespaces) else: @@ -244,11 +323,34 @@ class WikiqParser(): self.regex_revision_pairs = self.make_matchmake_pairs(regex_match_revision, regex_revision_label) self.regex_comment_pairs = self.make_matchmake_pairs(regex_match_comment, regex_comment_label) + if self.collapse_user is True: + if self.persist == PersistMethod.none: + revdata_type = RevDataCollapse + else: + revdata_type = RevDataCollapsePersistence + elif self.persist != PersistMethod.none: + revdata_type = RevDataPersistence + else: + revdata_type = RevDataBase + + regex_fields = [(field.name, list[str], dc.field(default=None)) for field in self.regex_schemas] + + self.revdata_type = make_dataclass('RevData_Parser', + fields=regex_fields, + bases=(revdata_type,)) + + self.revdata_type.pa_schema_fields = revdata_type.pa_schema_fields + self.regex_schemas + + + # print(list(map(lambda d: d.name, dc.fields(self.revdata_type)))) + # print(self.revdata_type.pa_schema_fields) if output_parquet is True: self.output_parquet = True self.pq_writer = None self.output_file = output_file + self.parquet_buffer = [] + self.parquet_buffer_size = parquet_buffer_size else: self.output_file = open(output_file,'w') @@ -273,7 +375,7 @@ class WikiqParser(): return rev_data def matchmake_revision(self, text, rev_data): - return self.matchmake_pairs(text, rev_data, self.regex_revision_pairs) + return self.matchmake_pairs(text, rev_data, self.regex_revision_pairs) def matchmake_comment(self, comment, rev_data): return self.matchmake_pairs(comment, rev_data, self.regex_comment_pairs) @@ -346,27 +448,18 @@ class WikiqParser(): # Iterate through a page's revisions for rev in page: - # initialize rev_data - rev_data = { - 'revid':rev.id, - 'date_time' : rev.timestamp.strftime('%Y-%m-%d %H:%M:%S'), - 'articleid' : page.id, - 'editorid' : "" if rev.deleted.user == True or rev.user.id is None else rev.user.id, - 'title' : '"' + page.title + '"', - 'namespace' : namespace, - 'deleted' : "TRUE" if rev.deleted.text else "FALSE" - } + rev_data = self.revdata_type(revid = rev.id, + date_time = datetime.fromtimestamp(rev.timestamp.unix()), + articleid = page.id, + editorid = "" if rev.deleted.user == True or rev.user.id is None else rev.user.id, + title = page.title, + deleted = rev.deleted.text, + namespace = namespace + ) rev_data = self.matchmake(rev, rev_data) - # if revisions are deleted, /many/ things will be missing - if rev.deleted.text: - rev_data['text_chars'] = "" - rev_data['sha1'] = "" - rev_data['revert'] = "" - rev_data['reverteds'] = "" - - else: + if not rev.deleted.text: # rev.text can be None if the page has no text if not rev.text: rev.text = "" @@ -378,33 +471,28 @@ class WikiqParser(): text_sha1 = sha1(bytes(rev.text, "utf8")).hexdigest() - rev_data['sha1'] = text_sha1 + rev_data.sha1 = text_sha1 # TODO rev.bytes doesn't work.. looks like a bug - rev_data['text_chars'] = len(rev.text) + rev_data.text_chars = len(rev.text) # generate revert data revert = rev_detector.process(text_sha1, rev.id) if revert: - rev_data['revert'] = "TRUE" - rev_data['reverteds'] = '"' + ",".join([str(x) for x in revert.reverteds]) + '"' + rev_data.revert = True + rev_data.reverteds = revert.reverteds else: - rev_data['revert'] = "FALSE" - rev_data['reverteds'] = "" + rev_data.revert = False # if the fact that the edit was minor can be hidden, this might be an issue - rev_data['minor'] = "TRUE" if rev.minor else "FALSE" + rev_data.minor = rev.minor if not rev.deleted.user: # wrap user-defined editors in quotes for fread - rev_data['editor'] = '"' + rev.user.text + '"' - rev_data['anon'] = "TRUE" if rev.user.id == None else "FALSE" + rev_data.editor = rev.user.text + rev_data.anon = rev.user.id == None - else: - rev_data['anon'] = "" - rev_data['editor'] = "" - #if re.match(r'^#redirect \[\[.*\]\]', rev.text, re.I): # redirect = True #else: @@ -414,13 +502,11 @@ class WikiqParser(): # if collapse user was on, lets run that if self.collapse_user: - rev_data['collapsed_revs'] = rev.collapsed_revs + rev_data.collapsed_revs = rev.collapsed_revs if self.persist != PersistMethod.none: - if rev.deleted.text: - for k in ["token_revs", "tokens_added", "tokens_removed", "tokens_window"]: - old_rev_data[k] = None - else: + + if not rev.deleted.text: if self.persist != PersistMethod.legacy: _, tokens_added, tokens_removed = state.update(rev.text, rev.id) @@ -435,12 +521,12 @@ class WikiqParser(): num_token_revs, num_tokens = calculate_persistence(old_tokens_added) - old_rev_data["token_revs"] = num_token_revs - old_rev_data["tokens_added"] = num_tokens - old_rev_data["tokens_removed"] = len(old_tokens_removed) - old_rev_data["tokens_window"] = PERSISTENCE_RADIUS-1 + rev_data.token_revs = num_token_revs + rev_data.tokens_added = num_tokens + rev_data.tokens_removed = len(old_tokens_removed) + rev_data.tokens_window = PERSISTENCE_RADIUS-1 - self.print_rev_data(old_rev_data) + self.print_rev_data(rev_data) else: self.print_rev_data(rev_data) @@ -457,10 +543,10 @@ class WikiqParser(): rev_id, rev_data, tokens_added, tokens_removed = item num_token_revs, num_tokens = calculate_persistence(tokens_added) - rev_data["token_revs"] = num_token_revs - rev_data["tokens_added"] = num_tokens - rev_data["tokens_removed"] = len(tokens_removed) - rev_data["tokens_window"] = len(window)-(i+1) + rev_data.token_revs = num_token_revs + rev_data.tokens_added = num_tokens + rev_data.tokens_removed = len(tokens_removed) + rev_data.tokens_window = len(window)-(i+1) self.print_rev_data(rev_data) page_count += 1 @@ -477,52 +563,50 @@ class WikiqParser(): def write_parquet_row(self, rev_data): - if 'deleted' in rev_data.keys(): - rev_data['deleted'] = True if rev_data['deleted'] == "TRUE" else False + padata = rev_data.to_pyarrow() + self.parquet_buffer.append(padata) - if 'minor' in rev_data.keys(): - rev_data['minor'] = True if rev_data['minor'] == "TRUE" else False + if len(self.parquet_buffer) >= self.parquet_buffer_size: + self.flush_parquet_buffer() - if 'anon' in rev_data.keys(): - rev_data['anon'] = True if rev_data['anon'] == "TRUE" else False + def flush_parquet_buffer(self): + schema = pa.schema(self.revdata_type.pa_schema_fields) + def row_to_col(rg, types): + cols = [] + first = rg[0] + for col in first: + cols.append([col]) - self.parquet_buffer.append(rev_data) + for row in rg[1:]: + for j in range(len(cols)): + cols[j].append(row[j]) - if len(self.parquet_buffer) >= self.parquet_buffer_size: - self.flush_parquet_buffer() + arrays = [] + for col, typ in zip(cols, types): + arrays.append(pa.array(col, typ)) + return arrays - def flush_parquet_buffer(self): - outtable = pd.DataFrame.from_records(self.parquet_buffer) - outtable = pa.Table.from_pandas(outtable) + outtable = pa.Table.from_arrays(row_to_col(self.parquet_buffer, schema.types), schema=schema) if self.pq_writer is None: - schema = outtable.schema - for regex_schema in self.regex_schemas: - schema.append(regex_schema) - self.pq_writer = pq.ParquetWriter(self.output_file, schema, flavor='spark') self.pq_writer.write_table(outtable) - + self.output_buffer = [] + def print_rev_data(self, rev_data): - if self.output_parquet is False: - printfunc = lambda rev_data: print("\t".join(rev_data), file=self.output_file) + printfunc = self.write_tsv_row else: printfunc = self.write_parquet_row - - # if it's the first time through, print the header - if self.urlencode: - for field in TO_ENCODE: - rev_data[field] = quote(str(rev_data[field])) - - if not self.printed_header: - printfunc(rev_data) - self.printed_header = True printfunc(rev_data) + def write_tsv_row(self, rev_data): + line = rev_data.to_tsv_line() + print(line, file=self.output_file) + def open_input_file(input_filename): if re.match(r'.*\.7z$', input_filename):