Datasets:
File size: 5,612 Bytes
146a370 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 | """Shared schemas and bounded-memory Parquet writing for dataset releases."""
import hashlib
import json
import logging
from pathlib import Path
import time
import pyarrow as pa
import pyarrow.parquet as pq
LOG = logging.getLogger(__name__)
RELEASE_FORMAT_VERSION = 2
# Input compatibility for releases exported before the public terminology change.
LEGACY_SCORE_COLUMNS = {
"negative_mining_scores": "negative_retrieval_scores",
"skipped_mining_scores": "skipped_retrieval_scores",
"mining_score": "retrieval_score",
}
def retrieval_score_schema(schema):
names = [LEGACY_SCORE_COLUMNS.get(name, name) for name in schema.names]
if len(names) != len(set(names)):
raise ValueError("Both legacy and retrieval score columns are present")
return pa.schema([field.with_name(name) for field, name in zip(schema, names)], metadata=schema.metadata)
GENERATION_ORIGIN = pa.struct([
("source_query_id", pa.string()),
("query_type", pa.string()),
("difficulty", pa.string()),
("character", pa.string()),
("query_length_target", pa.int64()),
("metadata_json", pa.string()),
])
OTHER_QUERY = pa.struct([
("query_id", pa.int64()),
("mining_query_id", pa.int64()),
("query", pa.string()),
("query_word_count", pa.int64()),
("generation_metadata", pa.list_(GENERATION_ORIGIN)),
])
TRAIN_SCHEMA = pa.schema([
("query_id", pa.int64()),
("mining_query_id", pa.int64()),
("query", pa.string()),
("query_word_count", pa.int64()),
("generation_metadata", pa.list_(GENERATION_ORIGIN)),
("other_queries", pa.list_(OTHER_QUERY)),
("positive_idx", pa.int64()),
("negative_idxs", pa.list_(pa.int64())),
("negative_retrieval_scores", pa.list_(pa.float64())),
("negative_reranker_scores", pa.list_(pa.float64())),
("negative_reranker_raw_scores", pa.list_(pa.float64())),
("positive_reranker_score", pa.float64()),
("positive_reranker_raw_score", pa.float64()),
("skipped_idxs", pa.list_(pa.int64())),
("skipped_retrieval_scores", pa.list_(pa.float64())),
])
REF_SCHEMA = pa.schema([
("finepdf_idx", pa.int64()), ("finepdf_id", pa.string()),
("source_shard", pa.string()), ("row_in_shard", pa.int64()),
])
NEGATIVE_ALIAS_SCHEMA = pa.schema([
("query_id", pa.int64()), ("mining_query_id", pa.int64()),
("positive_mining_id", pa.int64()), ("positive_idx", pa.int64()),
("removed_mining_id", pa.int64()), ("canonical_idx", pa.int64()),
("retained_mining_id", pa.int64()), ("reason", pa.string()),
("retrieval_score", pa.float64()), ("reranker_score", pa.float64()),
("reranker_raw_score", pa.float64()),
])
def write_json(path, value):
Path(path).write_text(json.dumps(value, indent=2, sort_keys=True) + "\n")
LOG.info("Wrote %s", path)
def sha256(path):
digest = hashlib.sha256()
with Path(path).open("rb") as stream:
for chunk in iter(lambda: stream.read(1024 * 1024), b""):
digest.update(chunk)
return digest.hexdigest()
class ShardedWriter:
"""Write fixed-size shards without buffering an entire output shard."""
def __init__(self, root, config, schema, rows_per_shard=100_000):
self.root = Path(root)
self.config = config
self.directory = self.root / config
self.directory.mkdir(parents=True)
self.schema = schema
self.capacity = rows_per_shard
self.writer = None
self.in_shard = 0
self.rows = 0
self.shards = []
self.started = time.monotonic()
self.last_report = self.started
def write(self, table):
table = table.cast(self.schema)
offset = 0
while offset < table.num_rows:
if self.writer is None:
self.path = self.directory / "part-{:05d}.parquet".format(len(self.shards))
self.writer = pq.ParquetWriter(self.path, self.schema, compression="zstd")
LOG.info("Opening output shard: %s", self.path)
count = min(self.capacity - self.in_shard, table.num_rows - offset)
self.writer.write_table(table.slice(offset, count), row_group_size=8192)
self.rows += count
self.in_shard += count
offset += count
if self.in_shard == self.capacity:
self._close_shard()
now = time.monotonic()
if now - self.last_report >= 60:
LOG.info("Writing %s: %s rows, %d completed shards, %.1f min elapsed",
self.config, format(self.rows, ",d"), len(self.shards), (now - self.started) / 60)
self.last_report = now
def _close_shard(self):
if self.writer is not None:
self.writer.close()
LOG.info("Closed %s (%s rows); computing SHA256", self.path, format(self.in_shard, ",d"))
self.shards.append({
"path": self.path.relative_to(self.root).as_posix(),
"start_row": self.rows - self.in_shard,
"num_rows": self.in_shard,
"sha256": sha256(self.path),
})
LOG.info("Output shard ready: %s (%.1f MiB; %s total %s rows)", self.path,
self.path.stat().st_size / (1024 * 1024), format(self.rows, ",d"), self.config)
self.writer = None
self.in_shard = 0
def close(self):
self._close_shard()
LOG.info("Finished %s: %s rows in %d shards (%.1f s)", self.config,
format(self.rows, ",d"), len(self.shards), time.monotonic() - self.started)
return {"num_rows": self.rows, "shards": self.shards}
|