X-Git-Url: https://code.communitydata.science/mediawiki_dump_tools.git/blobdiff_plain/02b3250a36a9a412ad5fe7f336caee861c2565b2..refs/heads/parquet_support:/wikiq diff --git a/wikiq b/wikiq index 7e75eda..75c1af8 100755 --- a/wikiq +++ b/wikiq @@ -8,6 +8,7 @@ import argparse import sys import os, os.path import re +from datetime import datetime,timezone from subprocess import Popen, PIPE from collections import deque @@ -24,6 +25,11 @@ PERSISTENCE_RADIUS=7 from deltas import SequenceMatcher from deltas import SegmentMatcher +import dataclasses as dc +from dataclasses import dataclass +import pyarrow as pa +import pyarrow.parquet as pq + class PersistMethod: none = 0 sequence = 1 @@ -34,7 +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 @@ -128,6 +133,11 @@ class WikiqPage(): return next(self.__revisions) +""" +A RegexPair is defined by a regular expression (pattern) and a label. +The pattern can include capture groups. If it does then each capture group will have a resulting column in the output. +If the pattern does not include a capture group, then only one output column will result. +""" class RegexPair(object): def __init__(self, pattern, label): self.pattern = re.compile(pattern) @@ -136,6 +146,15 @@ class RegexPair(object): if self.has_groups: self.capture_groups = list(self.pattern.groupindex.keys()) + def get_pyarrow_fields(self): + if self.has_groups: + fields = [pa.field(self._make_key(cap_group),pa.list_(pa.string())) + for cap_group in self.capture_groups] + else: + fields = [pa.field(self.label, pa.list_(pa.string()))] + + return fields + def _make_key(self, cap_group): return ("{}_{}".format(self.label, cap_group)) @@ -144,7 +163,7 @@ class RegexPair(object): temp_dict = {} # if there are named capture groups in the regex if self.has_groups: - # initialize the {capture_group_name:list} for each capture group + # if there are matches of some sort in this revision content, fill the lists for each cap_group if self.pattern.search(content) is not None: m = self.pattern.finditer(content) @@ -174,57 +193,249 @@ class RegexPair(object): # there are no capture groups, we just search for all the matches of the regex else: #given that there are matches to be made - if self.pattern.search(content) is not None: - m = self.pattern.findall(content) - temp_dict[self.label] = ', '.join(m) - else: - temp_dict[self.label] = None + if type(content) in(str, bytes): + if self.pattern.search(content) is not None: + m = self.pattern.findall(content) + temp_dict[self.label] = ', '.join(m) + else: + 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 +""" + +We used to use a dictionary to collect fields for the output. +Now we use dataclasses. Compared to a dictionary, this should help: +- prevent some bugs +- make it easier to output parquet data. +- use class attribute '.' syntax instead of dictionary syntax. +- improve support for tooling (autocomplete, type hints) +- use type information to define formatting rules + +Depending on the parameters passed into Wikiq, the output schema can be different. +Therefore, we need to end up constructing a dataclass with the correct output schema. +It also needs to have the correct pyarrow schema so we can write parquet files. + +The RevDataBase type has all the fields that will be output no matter how wikiq is invoked. +""" +@dataclass() +class RevDataBase(): + revid: int + date_time: datetime + articleid: int + editorid: int + title: str + namespace: int + deleted: bool + text_chars: int = None + revert: bool = None + reverteds: list[int] = None + sha1: str = None + minor: bool = None + editor: str = None + anon: bool = None + + # toggles url encoding. this isn't a dataclass field since it doesn't have a type annotation + urlencode = False + + # defines pyarrow schema. + # each field in the data class needs an entry in this array. + # the names should match and be in the same order. + # this isn't a dataclass field since it doesn't have a type annotation + 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(), nullable=True), + pa.field("title",pa.string()), + pa.field("namespace",pa.int32()), + pa.field("deleted",pa.bool_()), + pa.field("text_chars",pa.int32()), + pa.field("revert",pa.bool_(), nullable=True), + pa.field("reverteds",pa.list_(pa.int64()), nullable=True), + pa.field("sha1",pa.string()), + pa.field("minor",pa.bool_()), + pa.field("editor",pa.string()), + pa.field("anon",pa.bool_()) + ] + + # pyarrow is a columnar format, so most of the work happens in the flush_parquet_buffer function + def to_pyarrow(self): + return dc.astuple(self) + + # logic to convert each field into the wikiq tsv format goes here. + def to_tsv_row(self): + + row = [] + for f in dc.fields(self): + 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 self.urlencode and f.name in TO_ENCODE: + row.append(quote(str(s))) + else: + row.append(s) + + elif f.type == list[int]: + row.append('"' + ",".join([str(x) for x in val]) + '"') + + elif f.type == str: + if self.urlencode and f.name in TO_ENCODE: + row.append(quote(str(val))) + else: + row.append(val) + else: + row.append(val) + + return '\t'.join(map(str,row)) + + def header_row(self): + return '\t'.join(map(lambda f: f.name, dc.fields(self))) + +""" +If collapse=True we'll use a RevDataCollapse dataclass. +This class inherits from RevDataBase. This means that it has all the same fields and functions. + +It just adds a new field and updates the pyarrow schema. + +""" +@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] + +""" + +If persistence data is to be computed we'll need the fields added by RevDataPersistence. + +""" +@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 + +""" +class RevDataCollapsePersistence uses multiple inheritence to make a class that has both persistence and collapse 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): + 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): """ Parameters: persist : what persistence method to use. Takes a PersistMethod value """ self.input_file = input_file - self.output_file = output_file + self.collapse_user = collapse_user self.persist = persist - self.printed_header = False self.namespaces = [] self.urlencode = urlencode self.revert_radius = revert_radius - + if namespaces is not None: self.namespace_filter = set(namespaces) else: self.namespace_filter = None + self.regex_schemas = [] 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) + + + # This is where we set the type for revdata. + + 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 + + # if there are regex fields, we need to add them to the revdata type. + regex_fields = [(field.name, list[str], dc.field(default=None)) for field in self.regex_schemas] + + # make_dataclass is a function that defines a new dataclass type. + # here we extend the type we have already chosen and add the regular expression types + self.revdata_type = dc.make_dataclass('RevData_Parser', + fields=regex_fields, + bases=(revdata_type,)) + # we also need to make sure that we have the right pyarrow schema + self.revdata_type.pa_schema_fields = revdata_type.pa_schema_fields + self.regex_schemas + + self.revdata_type.urlencode = self.urlencode + + self.schema = pa.schema(self.revdata_type.pa_schema_fields) + + # here we initialize the variables we need for output. + 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.print_header = True + if output_file == sys.stdout: + + self.output_file = output_file + else: + self.output_file = open(output_file,'w') + self.output_parquet = False def make_matchmake_pairs(self, patterns, labels): if (patterns is not None and labels is not None) and \ (len(patterns) == len(labels)): - return [RegexPair(pattern, label) for pattern, label in zip(patterns, labels)] + result = [] + for pattern, label in zip(patterns, labels): + rp = RegexPair(pattern, label) + result.append(rp) + self.regex_schemas = self.regex_schemas + rp.get_pyarrow_fields() + return result elif (patterns is None and labels is None): return [] else: sys.exit('Each regular expression *must* come with a corresponding label and vice versa.') - def matchmake(self, rev, rev_data): - rev_data = self.matchmake_revision(rev.text, rev_data) + def matchmake_revision(self, rev, rev_data): + rev_data = self.matchmake_text(rev.text, rev_data) rev_data = self.matchmake_comment(rev.comment, rev_data) return rev_data - def matchmake_revision(self, text, rev_data): - return self.matchmake_pairs(text, rev_data, self.regex_revision_pairs) + def matchmake_text(self, text, rev_data): + 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) @@ -280,7 +491,7 @@ class WikiqParser(): if self.persist != PersistMethod.none: window = deque(maxlen=PERSISTENCE_RADIUS) - + if self.persist == PersistMethod.sequence: state = mwpersistence.DiffState(SequenceMatcher(tokenizer = wikitext_split), revert_radius=PERSISTENCE_RADIUS) @@ -297,27 +508,19 @@ 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, - 'editor_id' : "" 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.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: + # create a new data object instead of a dictionary. + rev_data = self.revdata_type(revid = rev.id, + date_time = datetime.fromtimestamp(rev.timestamp.unix(), tz=timezone.utc), + 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_revision(rev, rev_data) + + if not rev.deleted.text: # rev.text can be None if the page has no text if not rev.text: rev.text = "" @@ -326,36 +529,30 @@ class WikiqParser(): if rev.sha1: text_sha1 = rev.sha1 else: - 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" - - else: - rev_data['anon'] = "" - rev_data['editor'] = "" - + rev_data.editor = rev.user.text + rev_data.anon = rev.user.id is None + #if re.match(r'^#redirect \[\[.*\]\]', rev.text, re.I): # redirect = True #else: @@ -365,13 +562,11 @@ 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 + # get the 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) @@ -386,10 +581,10 @@ 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 + 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 self.print_rev_data(old_rev_data) @@ -408,11 +603,10 @@ 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 @@ -420,22 +614,79 @@ class WikiqParser(): print("Done: %s revisions and %s pages." % (rev_count, page_count), file=sys.stderr) + # remember to flush the parquet_buffer if we're done + if self.output_parquet is True: + self.flush_parquet_buffer() + self.pq_writer.close() + + else: + self.output_file.close() + + + """ + For performance reasons it's better to write parquet in batches instead of one row at a time. + So this function just puts the data on a buffer. If the buffer is full, then it gets flushed (written). + """ + def write_parquet_row(self, rev_data): + padata = rev_data.to_pyarrow() + self.parquet_buffer.append(padata) + + if len(self.parquet_buffer) >= self.parquet_buffer_size: + self.flush_parquet_buffer() + + + """ + Function that actually writes data to the parquet file. + It needs to transpose the data from row-by-row to column-by-column + """ + def flush_parquet_buffer(self): + + """ + Returns the pyarrow table that we'll write + """ + def rows_to_table(rg, schema): + cols = [] + first = rg[0] + for col in first: + cols.append([col]) + + for row in rg[1:]: + for j in range(len(cols)): + cols[j].append(row[j]) + + arrays = [] + for col, typ in zip(cols, schema.types): + arrays.append(pa.array(col, typ)) + return pa.Table.from_arrays(arrays, schema=schema) + + outtable = rows_to_table(self.parquet_buffer, self.schema) + if self.pq_writer is None: + self.pq_writer = pq.ParquetWriter(self.output_file, schema, flavor='spark') + + self.pq_writer.write_table(outtable) + self.parquet_buffer = [] + + # depending on if we are configured to write tsv or parquet, we'll call a different function. def print_rev_data(self, rev_data): - # 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: - print("\t".join([str(k) for k in sorted(rev_data.keys())]), file=self.output_file) - self.printed_header = True + if self.output_parquet is False: + printfunc = self.write_tsv_row + else: + printfunc = self.write_parquet_row - print("\t".join([str(v) for k, v in sorted(rev_data.items())]), file=self.output_file) + printfunc(rev_data) + + def write_tsv_row(self, rev_data): + if self.print_header: + print(rev_data.header_row(), file=self.output_file) + self.print_header = False + + line = rev_data.to_tsv_row() + print(line, file=self.output_file) def open_input_file(input_filename): if re.match(r'.*\.7z$', input_filename): - cmd = ["7za", "x", "-so", input_filename, '*'] + cmd = ["7za", "x", "-so", input_filename, "*.xml"] elif re.match(r'.*\.gz$', input_filename): cmd = ["zcat", input_filename] elif re.match(r'.*\.bz2$', input_filename): @@ -448,13 +699,19 @@ def open_input_file(input_filename): return input_file -def open_output_file(input_filename): - # create a regex that creates the output filename +def get_output_filename(input_filename, parquet = False): output_filename = re.sub(r'\.(7z|gz|bz2)?$', '', input_filename) output_filename = re.sub(r'\.xml', '', output_filename) - output_filename = output_filename + ".tsv" - output_file = open(output_filename, "w") + if parquet is False: + output_filename = output_filename + ".tsv" + else: + output_filename = output_filename + ".parquet" + return output_filename +def open_output_file(input_filename): + # create a regex that creates the output filename + output_filename = get_output_filename(input_filename, parquet = False) + output_file = open(output_filename, "w") return output_file parser = argparse.ArgumentParser(description='Parse MediaWiki XML database dumps into tab delimitted data.') @@ -464,7 +721,7 @@ parser.add_argument('dumpfiles', metavar="DUMPFILE", nargs="*", type=str, help="Filename of the compressed or uncompressed XML database dump. If absent, we'll look for content on stdin and output on stdout.") parser.add_argument('-o', '--output-dir', metavar='DIR', dest='output_dir', type=str, nargs=1, - help="Directory for output files.") + help="Directory for output files. If it ends with .parquet output will be in parquet format.") parser.add_argument('-s', '--stdout', dest="stdout", action="store_true", help="Write output to standard out (do not create dump file)") @@ -503,6 +760,8 @@ parser.add_argument('-CPl', '--comment-pattern-label', dest="regex_comment_label args = parser.parse_args() + + # set persistence method if args.persist is None: @@ -520,6 +779,7 @@ else: namespaces = None if len(args.dumpfiles) > 0: + output_parquet = False for filename in args.dumpfiles: input_file = open_input_file(filename) @@ -529,13 +789,16 @@ if len(args.dumpfiles) > 0: else: output_dir = "." + if output_dir.endswith(".parquet"): + output_parquet = True + print("Processing file: %s" % filename, file=sys.stderr) if args.stdout: output_file = sys.stdout else: filename = os.path.join(output_dir, os.path.basename(filename)) - output_file = open_output_file(filename) + output_file = get_output_filename(filename, parquet = output_parquet) wikiq = WikiqParser(input_file, output_file, @@ -547,13 +810,14 @@ if len(args.dumpfiles) > 0: regex_match_revision = args.regex_match_revision, regex_revision_label = args.regex_revision_label, regex_match_comment = args.regex_match_comment, - regex_comment_label = args.regex_comment_label) + regex_comment_label = args.regex_comment_label, + output_parquet=output_parquet) wikiq.process() # close things input_file.close() - output_file.close() + else: wikiq = WikiqParser(sys.stdin, sys.stdout,