A Python tool for deep inspection and analysis of Apache Parquet files, providing detailed insights into file structure, metadata, and binary layout.
For an example interactive HTML report generated by this tool, see: https://clee704.github.io/parquet-analyzer/examples/example.html.
pip install parquet-analyzer- Python 3.11+
parquet-analyzer ships two complementary CLI surfaces:
- Verb-noun subcommands (
parquet-analyzer file kv ...,column show ..., etc.) — small, composable, AI-friendly. One subcommand → one JSON object. Footer-only and fast. See Subcommands. - Whole-file modes (
parquet-analyzer <path> [--output-mode ...]) — the original "show me everything" surface for interactive exploration and the HTML report.
# Verb-noun subcommands (footer-only, fast)
parquet-analyzer file summary example.parquet
parquet-analyzer column show example.parquet --column ints
parquet-analyzer rowgroup list example.parquet
# Whole-file: analyze and emit the JSON summary/footer/pages bundle
parquet-analyzer example.parquet
# Show raw segment structures (offsets, lengths, thrift payloads)
parquet-analyzer --output-mode segments example.parquet
# Generate an interactive HTML report and save it to disk
parquet-analyzer --output-mode html -o report.html example.parquet
# Generate an HTML report with selected sections only
parquet-analyzer --output-mode html \
--html-sections summary schema key-value-metadata row-groups columns segments \
-o report.html example.parquet
# Enable debug logging while running any mode
parquet-analyzer --log-level DEBUG example.parquet
# Run via python -m if the console script is unavailable
python -m parquet_analyzer example.parquetThe verb-noun surface answers one question per invocation and is built for
chaining with jq and consumption by AI agents. The eight verb-noun
subcommands under file/rowgroup/column are footer-only — they do not
walk page headers or read page bodies. A ninth verb,
show, offers path-addressed tree
navigation (also bounded — it never walks page headers unless you pass
--walk-pages). The page verb is the escape
hatch into the page layer: it lists a chunk's pages, dumps a single page
header, extracts raw body bytes, and decodes a page body into its
encoding-faithful structure. Reading page bodies is opt-in by construction —
you pay the per-page cost only for the page you name.
- One subcommand → one JSON object on stdout. List-shaped outputs wrap in
{items, total, returned, truncated}. - Every output object carries a
$schemafield of the formparquet-analyzer/v1/<command>so downstream tools can validate the shape. --schema-versionon any subcommand short-circuits to print just the schema URI — no file is opened.--limit Ncaps list-shaped outputs. Truncation is reported explicitly via thetruncatedfield; nothing is silently dropped.-o / --output PATHwrites JSON to a file instead of stdout.--log-level {DEBUG,INFO,WARNING,ERROR,CRITICAL}is the top-level global option.- Navigation cross-link. Every
rowgroup/columnlistandshowitem carries a_path— its canonicalshownavigation path (row_groups/2/columns/1). Use the flat commands as a search/sort surface, then feed a_pathstraight back intoshowto drill in (no need to hand-build the path):parquet-analyzer column list data.parquet | jq -r '.items | max_by(.compressed_size)._path' \ | xargs parquet-analyzer show data.parquet
- Errors → stderr as a single JSON line:
{"$schema": "parquet-analyzer/v1/error", "error": "<code>", "message": "<human>", "fix": "<retry command>"}. The process exits non-zero on operational errors;file validatereports parse failures as findings (exit 0). - Column path matching. Every emitted column carries both
column(dot-joined display string) andpath(list of segments).--column NAMEmatches against the dot-joined form. Ambiguous matches (when multiple distinct paths join to the same string) produce a JSON error listing the candidate paths.
Design contract for what goes in the output: see
docs/output-principles.md. The TL;DR is "footer-bounded and walk-free": every emitted field is derivable from the parsed footer plus, at most, one extra parse of an index thrift the writer already wrote (OffsetIndex, ColumnIndex, BloomFilter header). Anything that would require walking per-page thrift headers or reading page bodies is handled by thepagesubcommand surface — the deliberate, per-page opt-in into header walks and body reads.
$ parquet-analyzer file summary example.parquet
{
"$schema": "parquet-analyzer/v1/file-summary",
"num_rows": 891,
"num_row_groups": 1,
"num_columns": 12,
"uncompressed_page_size": 90429,
"compressed_page_size": 38839,
"column_index_size": 0,
"offset_index_size": 0,
"bloom_filter_size": 0,
"footer_size": 1162,
"file_size": 40013,
"footer_offset": 38843
}
$ parquet-analyzer file kv example.parquet
$ parquet-analyzer file kv example.parquet --key ARROW:schema # filter by key
{
"$schema": "parquet-analyzer/v1/file-kv",
"items": [{"key": "ARROW:schema", "value": "..."}],
"returned": 1, "total": 1, "truncated": false,
"filter_key": "ARROW:schema"
}
$ parquet-analyzer file schema example.parquet
{
"$schema": "parquet-analyzer/v1/file-schema",
"elements": [
{"name": "root", "repetition_type": "REQUIRED", "num_children": 12},
{"name": "PassengerId", "type": "INT64", ...},
...
]
}
$ parquet-analyzer file validate example.parquet
{
"$schema": "parquet-analyzer/v1/file-validate",
"path": "example.parquet",
"valid": true,
"errors": []
}file kv preserves duplicate keys and original ordering (parquet's spec
allows duplicate KV entries). file validate runs footer-only structural
checks: magic-number parse, row-count consistency between file and row
groups, per-row-group column count matches schema, and column chunks with
rows have non-zero compressed size. Parse failures (bad magic, unparseable
footer) are reported as findings with valid: false rather than as CLI
errors. Deep validation (byte-range overlap detection, page-count vs.
column metadata) requires an eager walk and will land in a future release.
$ parquet-analyzer rowgroup list example.parquet
{
"$schema": "parquet-analyzer/v1/rowgroup-list",
"items": [
{"row_group": 0, "num_rows": 891, "file_offset": 4,
"total_byte_size": 306419, "total_compressed_size": 38839,
"num_columns": 12}
],
"returned": 1, "total": 1, "truncated": false
}
$ parquet-analyzer rowgroup show example.parquet --row-group 0
{
"$schema": "parquet-analyzer/v1/rowgroup-show",
"row_group": 0,
"num_rows": 891,
"file_offset": 4,
"total_byte_size": 306419,
"total_compressed_size": 38839,
"num_columns": 12,
"columns": [
{
"row_group": 0, "column": "PassengerId", "path": ["PassengerId"],
"type": "INT64", "encodings": ["PLAIN"], "codec": "SNAPPY",
"num_values": 891, "compressed_size": 4357, "uncompressed_size": 7155,
"chunk_offset": 4, "chunk_length": 4357,
"data_page_offset": 4, "dictionary_page_offset": null,
"has_dictionary": false,
"has_offset_index": false, "offset_index_offset": null, "offset_index_length": null,
"has_column_index": false, "column_index_offset": null, "column_index_length": null,
"has_bloom_filter": false, "bloom_filter_offset": null, "bloom_filter_length": null,
"num_pages": null, "num_pages_known": false,
"num_pages_hint": "no OffsetIndex; re-run with --walk-pages to count pages (reads page headers)",
...
},
...
]
}# One entry per (row_group, column_chunk); --row-group N to filter
$ parquet-analyzer column list example.parquet
$ parquet-analyzer column list example.parquet --row-group 0 --limit 5
# Aggregated view across row groups for one named column
$ parquet-analyzer column show example.parquet --column Sex
{
"$schema": "parquet-analyzer/v1/column-show",
"column": "Sex",
"path": ["Sex"],
"type": "BYTE_ARRAY",
"num_row_groups": 1,
"total_num_values": 891,
"total_compressed_size": 501,
"total_uncompressed_size": 897,
"row_groups": [
{
"row_group": 0, "column": "Sex", "path": ["Sex"],
"encodings": ["PLAIN", "RLE_DICTIONARY"], "codec": "SNAPPY",
"num_values": 891, "compressed_size": 501, "uncompressed_size": 897,
"chunk_offset": 24256, "chunk_length": 501,
"data_page_offset": 24276, "dictionary_page_offset": 24256,
"has_dictionary": true,
"has_offset_index": false, "offset_index_offset": null, "offset_index_length": null,
"has_column_index": false, "column_index_offset": null, "column_index_length": null,
"has_bloom_filter": false, "bloom_filter_offset": null, "bloom_filter_length": null,
"num_pages": null, "num_pages_known": false,
"num_pages_hint": "no OffsetIndex; re-run with --walk-pages to count pages (reads page headers)",
"statistics": {"min": "...", "max": "...", "null_count": 0}
}
]
}num_pages is reported (num_pages_known: true) when the chunk has an
OffsetIndex — an O(1) lookup. Without one, counting pages requires walking
page headers, which the footer-bounded default avoids (num_pages: null,
num_pages_known: false, plus a num_pages_hint string pointing at
--walk-pages so the output is self-describing). Pass --walk-pages (on
column show, column list, or rowgroup show) to opt into that walk and
populate num_pages anyway — it reads the page headers of every selected
chunk, so --limit caps the output but not the walk. SNPW (Spark Native
Parquet Writer) writes OffsetIndex on every chunk; pyarrow does when
write_page_index=True; older parquet-mr and many DuckDB files do not.
Byte-range pairs. Every per-chunk output carries the seek-and-read
pair chunk_offset / chunk_length, plus offset_index_offset /
offset_index_length, column_index_offset / column_index_length,
and bloom_filter_offset / bloom_filter_length (each null when the
writer didn't emit the corresponding structure). chunk_offset is the
dictionary page offset when a dictionary is present, otherwise the data
page offset (parquet's spec: dictionary always precedes data within a
chunk; total_compressed_size already includes the dictionary). Together
these let an AI agent issue read(chunk_offset, chunk_length) against
the file and get the entire compressed chunk bytes without re-parsing the
footer.
show FILE [PATH] renders the node at PATH plus its immediate children
as stubs, each annotated with the canonical path to descend into it. You
explore the file like a map, one bounded step at a time, by extending the
path along the row_groups → columns → pages spine:
parquet-analyzer show data.parquet # file: row groups (stubs)
parquet-analyzer show data.parquet row_groups/0 # that group: columns (stubs)
parquet-analyzer show data.parquet row_groups/0/columns/3 # column metadata + stats; pages (stubs)
parquet-analyzer show data.parquet row_groups/0/columns/3/pages/5 # that page (direct-seek)Paths are canonical and index-based. Each child stub carries a _path
to feed back into the next show, columns also carry their display name,
and a _navigation block reports the current path, its parent, and the
node kind — so a caller (human or agent) never has to construct a path by
hand:
(The show envelope is a v1 CLI response; the node stubs inside it carry the
v3 tree-node _location address — {offset, length}, the file byte range.)
Listing a column's pages never forces a page-header walk. With an
OffsetIndex the pages are listed from it (one small read, independent of
page count). Without one, the listing is withheld behind an explicit
--walk-pages opt-in: pages is null and the _navigation block carries
a walk_required affordance instead of paying an O(pages) walk:
"pages": null,
"_navigation": {"path": "...", "parent": "...", "kind": "column_chunk",
"children_total": null, "children_shown": 0, "children_truncated": false,
"walk_required": true, "reason": "no OffsetIndex",
"hint": "re-run with '.../pages/<n> --walk-pages' ..."}--walk-pages enables listing/addressing pages of such a column (it reads
every page header). A column can have many thousands of pages, so the child
listing is capped by --limit N (default 100; 0 lists all); the
_navigation block reports children_total / children_shown /
children_truncated, and truncation only bounds the listing — every child
stays addressable by its index. show is tree-structured; the verb-noun
subcommands above remain the way to get flat, aggregated views.
The page verb is the escape hatch into the page layer — the one surface that
deliberately reads past the footer. Where the footer-only verbs stop at the
column chunk, page walks into the pages of a single chunk to list them, dump
a header, extract raw body bytes, or decode a body into the structure of its
actual on-disk encoding. You pay the page cost only for the page(s) you name.
All four nouns share a page selector: --column NAME, --page-index N (over
the chunk's full page order — the dictionary page is index 0 when present,
then the data pages; negative indexes count from the end), and --row-group M
(required only when the file has more than one row group). The singular verbs
report both page_index (full order) and data_page_index (position among
data pages only — the OffsetIndex correspondence, null for the dictionary
page).
Equivalently, any page noun accepts a navpath positional — the same
_path that page list (and show) emit — instead of the selectors, so a
page list → page decode round-trip is copy-paste. For the singular verbs
the navpath addresses a page (row_groups/0/columns/4/pages/1); for page list it scopes to a row group (row_groups/0) or a column chunk
(row_groups/0/columns/4). The navpath and the --column/--page-index/
--row-group selectors are mutually exclusive. Every singular-verb output
echoes the resolved _path.
--limit N bounds the potentially-large output (page list and page decode accept it; page header and page extract don't — they return a
single object or raw bytes). For page list it caps the items. For page decode it bounds each collection in the view to the first N rows of the
page: each level stream's levels, the resolved/PLAIN values, and the
dictionary index runs — with the runs clipped so they cover exactly N
rows (a single {value: 0, length: 1000000} run shows as {value: 0, length: 5} under --limit 5, not in full). Every bounded collection reports its own
total / returned / truncated, so a truncation is always explicit.
--kind statistics ignores it (the header's statistics aren't per-row).
Default is no limit (the whole page).
One subtlety on nullable columns: the level streams have one entry per row
(nulls included), while the value/index collections — resolved values, PLAIN
values, and the dictionary runs — only carry the present (non-null)
values, since parquet stores no index or value for a null. So --limit N
bounds the level streams to the first N rows and the value/index collections to
the first N present values; the two coincide exactly when the page has no
nulls, and otherwise their totals differ (rows vs. non-null count) and their
windows diverge by the nulls in the prefix.
This makes --limit the tool for a page that's tiny on disk but expands to
millions of values (one long RLE run): page decode … --kind values --limit 10
samples the first ten decoded values without materializing the rest.
$ parquet-analyzer page list data.parquet --column Sex | jq -r '.items[1]._path' \
| xargs parquet-analyzer page decode data.parquetpage list is the page-walk surface: a column's (or every column's) pages
as lightweight stubs. It is cheap when the writer emitted an OffsetIndex (the
extents come from it, and stubs carry first_row_index); otherwise it walks
the per-page headers. Like the other list verbs it wraps in
{items, total, returned, truncated} and every item carries a _path.
$ parquet-analyzer page list example.parquet --column Sex
{
"$schema": "parquet-analyzer/v1/page-list",
"items": [
{"row_group": 0, "column": "Sex", "_path": "row_groups/0/columns/4/pages/0",
"page_index": 0, "data_page_index": null, "kind": "dictionary_page",
"offset": 24256, "length": 33, "first_row_index": null},
{"row_group": 0, "column": "Sex", "_path": "row_groups/0/columns/4/pages/1",
"page_index": 1, "data_page_index": 0, "kind": "data_page",
"offset": 24289, "length": 468, "first_row_index": null}
],
"returned": 2, "total": 2, "truncated": false, "column": "Sex"
}page header dumps one page's full header fields. It is version-aware: V1
data pages report the level encodings, V2 data pages report
num_nulls/num_rows/is_compressed and the level byte lengths.
$ parquet-analyzer page header example.parquet --column Sex --page-index 1
{
"$schema": "parquet-analyzer/v1/page-header",
"_path": "row_groups/0/columns/4/pages/1",
"row_group": 0, "column": "Sex", "page_index": 1, "data_page_index": 0,
"kind": "data_page", "page_type": "DATA_PAGE", "offset": 24289,
"header_size": 20, "compressed_size": 448, "uncompressed_size": 846,
"num_values": 891, "encoding": "RLE_DICTIONARY",
"definition_level_encoding": "RLE", "repetition_level_encoding": "RLE",
"statistics": null
}page extract emits one page's raw body bytes. --decompress removes the
page codec (page-type aware: a V2 page keeps its uncompressed level streams and
expands only the values section). --as {hex,base64,raw} chooses the encoding;
--as raw is the JSON escape hatch — it writes the bytes verbatim to stdout
(or -o FILE) with no JSON envelope.
$ parquet-analyzer page extract example.parquet --column Sex --page-index 0 --as hex
$ parquet-analyzer page extract example.parquet --column Sex --page-index 1 --decompress --as raw -o body.binpage decode decodes a page body. With no --kind it emits the full
encoding-faithful decode in one object — the definition/repetition level
streams plus the values section in its native encoding form, which the page's
encoding determines (no choice to make):
$ parquet-analyzer page decode example.parquet row_groups/0/columns/4/pages/1 --limit 3
{
"$schema": "parquet-analyzer/v1/page-decode",
"_path": "row_groups/0/columns/4/pages/1",
"row_group": 0, "column": "Sex", "page_index": 1, "data_page_index": 0,
"encoding": "RLE_DICTIONARY", "num_values": 891, "num_nulls": 0,
"definition_levels": {"bit_width": 1, "total": 891, "returned": 3, "truncated": true, "levels": [1, 1, 1]},
"repetition_levels": null,
"encoded_values": {
"kind": "dictionary_indices", "bit_width": 2, "total": 891, "returned": 3, "truncated": true,
"runs": [{"kind": "rle", "value": 0, "length": 1}, {"kind": "rle", "value": 1, "length": 2}]
}
}encoded_values is the values section as it is literally stored: for a
dictionary-encoded page the index RLE/bit-packed runs (kind: "dictionary_indices"); for a PLAIN page the verbatim values (kind: "plain").
This is the encoding-faithful view — the data the encoding scheme actually
stores, not a flattened logical reconstruction.
--kind narrows to a single view (and is the only way to get the resolved
physical values):
values— the page's resolved physical values, with dictionary indices resolved through the sibling dictionary page to the physical-type values. Nulls are skipped, sototalis the non-null count.levels— just the definition and repetition level streams (each abit_widthplus the expandedlevels; eithernullwhen the column has no such block — a flat column has no repetition levels, a required column no definition levels).rle-runs— just the dictionary index runs (kind_not_availableon a PLAIN page).statistics— the page header's statistics (header-only; no body decode).
$ parquet-analyzer page decode example.parquet --column Sex --page-index 1 --kind values --limit 2A page whose encoding or codec this tool does not yet decode returns a
structured error (encoding_not_supported / codec_not_supported) that names
the unsupported scheme — fall back to page extract for the raw bytes.
Every subcommand accepts --schema-version, which short-circuits to print
the JSON $schema URI for that subcommand's output without opening the
file. Useful for tools that want to fetch the matching schema definition.
$ parquet-analyzer column show --schema-version
{
"$schema": "parquet-analyzer/v1/column-show"
}Errors are written to stderr as a single JSON line with the contract:
{
"$schema": "parquet-analyzer/v1/error",
"error": "column_not_found",
"message": "column 'foo' not found. Available: a, b, c",
"fix": "parquet-analyzer column list example.parquet"
}The fix field always contains an exact command to retry. The process
exits non-zero on operational errors. file validate is the exception:
parse failures are reported as findings (valid: false, exit 0) so the
subcommand stays useful for "is this file partially-written?" queries.
The default output provides a structured JSON payload with three main sections:
{
"summary": {
"num_rows": 10,
"num_row_groups": 1,
"num_columns": 2,
"num_pages": 2,
"num_data_pages": 2,
"num_v1_data_pages": 2,
"num_v2_data_pages": 0,
"num_dict_pages": 0,
"page_header_size": 47,
"uncompressed_page_data_size": 130,
"compressed_page_data_size": 96,
"uncompressed_page_size": 177,
"compressed_page_size": 143,
"column_index_size": 48,
"offset_index_size": 23,
"bloom_filter_size": 0,
"footer_size": 527,
"file_size": 753
}
}Complete Parquet file metadata including:
- Schema definition with column types and repetition levels
- Row group information
- Column chunk metadata
- Encoding and compression details
Detailed breakdown of all pages organized by column:
- Data pages with encoding and statistics
- Dictionary pages
- Column indexes
- Offset indexes
- Bloom filters
When using --output-mode segments, the tool outputs a detailed segment-by-segment breakdown showing:
[
{
"offset": 0,
"length": 4,
"name": "magic_number",
"value": "PAR1"
},
{
"offset": 4,
"length": 24,
"name": "page",
"value": [
{
"offset": 5,
"length": 1,
"name": "type",
"value": 0,
"metadata": {
"type": "i32",
"enum_type": "PageType",
"enum_name": "DATA_PAGE"
}
}
]
}
]This mode is useful for:
- Understanding exact binary layout
- Analyzing file format compliance
- Optimizing file structure
Emits a standalone HTML document with collapsible sections for summary statistics, schema, key-value metadata, row groups, aggregated column statistics, segments, and the raw footer. Use the --html-sections flag to control which sections are rendered:
parquet-analyzer --output-mode html \
--html-sections summary schema key-value-metadata row-groups columns segments \
-o report.html \
example.parquetExample: https://clee704.github.io/parquet-analyzer/examples/example.html
parquet-analyzer is also a Python library. Anything the CLI does is available as an importable function, so you can build encoding-level verification scripts without shelling out.
from parquet_analyzer import ParquetFile
with ParquetFile("example.parquet") as pf:
# Footer-only — instant on any file size.
print(pf.num_rows, pf.num_row_groups, pf.num_columns)
print(pf.schema)
print(pf.kv_metadata_lookup("com.acme.author"))
# Walk row groups + column chunks — still footer-only, no body reads.
for rg in pf.row_groups:
for cc in rg.columns:
print(cc.path, cc.type, cc.encodings, cc.codec, cc.num_values)ParquetFile reads only the footer on construction (a few KB at end of file). Per-row-group / per-column / per-page metadata is exposed through wrapper classes that defer expensive parsing until you actually ask for it.
Two summary surfaces:
pf.footer_summary— cheap, footer-only (row/group/column counts, footer + file size, aggregate compressed/uncompressed column-chunk sizes).pf.full_summary— same shape as legacyget_summary(), includes per-page counts (num_pages,num_data_pages, etc.). Triggers a full eager walk.
Per-chunk lazy page walking:
# How many pages does this chunk have? Fast (O(1)) if the writer included
# an OffsetIndex; falls back to walking otherwise.
cc = pf.row_groups[0].columns[0]
if cc.has_offset_index:
print(f"chunk has {cc.num_pages} pages (cheap lookup via OffsetIndex)")
else:
print(f"chunk has {cc.num_pages} pages (paid full walk to count them)")
# Iterate the page headers themselves (walks once, caches).
for page in cc.pages():
print(page.type, page.encoding, page.num_values, page.offset)columnchunk.pages() walks only that one chunk's page headers (cached after first call) — much cheaper than the full-file walk. num_pages is even cheaper when an OffsetIndex is present (SNPW always writes one; pyarrow does when write_page_index=True; older parquet-mr files often don't).
Each Page decodes its own body on demand — faithfully to the encoding,
not as a flattened reconstruction — without walking the rest of the file.
Every encoded stream is exposed in its own structure:
cc = pf.row_groups[0].columns[0]
page = cc.page(1) # seeks via the OffsetIndex when present
decoded = page.decode() # a DecodedPage (cached on the page)
decoded.num_nulls # V2: from the header; V1: from def levels
# Level streams: RleBitPackedStream (the encoding parquet uses for levels),
# or None when the column has no such block on disk (required / non-repeated).
decoded.definition_levels # RleBitPackedStream | None
decoded.repetition_levels # RleBitPackedStream | None
lvl = decoded.definition_levels
lvl.bit_width # packing width (from max_definition_level)
lvl.runs # ordered (RleRun | BitPackedRun) — on-disk structure
lvl.values # the expanded per-value levels
# Values section: PlainValues for PLAIN, or — because dictionary indices use
# the SAME RLE/bit-packed encoding as levels — an RleBitPackedStream of the
# raw indices for a dictionary page.
decoded.values # PlainValues | RleBitPackedStream
# e.g. for a dict page: decoded.values.bit_width / .runs / .values (indices)
# Physical-TYPE values (one level below logical): the encoding decoded and
# dict indices resolved, but NOT logical-type-interpreted — a UTF8 column
# yields b'S', not 'S' (a logical_values() companion is planned):
page.physical_values()
page.definition_levels() # convenience: expanded levels ([0]*n if no block)
page.repetition_levels()
page.raw_body() # the on-disk body bytes (no decode)
# The chunk's dictionary (decoded once, cached) when it has a dictionary page:
cc.dictionary() # list of physical-type values, or None
cc.max_definition_level # schema-derived, cheap
cc.max_repetition_leveldecode() gives you the encoding's own data: an RleBitPackedStream
(bit_width + ordered RleRun / BitPackedRun) for levels and dictionary
indices, and PlainValues for PLAIN. The values section carries only the
non-null entries (num_values - num_nulls); the nulls live in the
definition levels — this is the on-disk shape, not a reassembled column.
There are three levels between the on-disk bytes and the logical values:
the encoding representation (decode().values — indices/runs/PLAIN
bytes), the physical-type values (physical_values() — decoded to the
parquet physical type with the dictionary resolved, e.g. b'S'), and the
logical-type values (the physical values reinterpreted via the column's
logical type, e.g. 'S' / Decimal / a timestamp — not yet implemented).
The dictionary is an encoding, not a type, so physical_values() yields
the same physical type whether or not the page was dictionary-encoded.
V1 (length-prefixed levels inside the compressed body) and V2 (uncompressed
levels ahead of an optionally-compressed values section) are handled
transparently. An out-of-scope encoding or an undecompressable codec on
decode(), a non-data page on decode(), or a missing dictionary on
physical_values() raises a typed PageDecodeError subclass
(UnsupportedEncodingError,
UnsupportedCodecError, UnsupportedPageTypeError, MissingDictionaryError)
carrying a stable .code.
When you genuinely need the complete byte-range view of every structural element:
with ParquetFile("example.parquet") as pf:
every_segment = pf.all_segments() # sorted + gap-filled
every_page = pf.all_pages() # per-column pages tree
chunk_offset_map = pf.column_offset_map # legacy parse_parquet_file()[1] shapeThese are the only methods that trigger the full per-page Thrift walk that the legacy parse_parquet_file() always did. The CLI's --output-mode segments mode uses pf.all_segments(); the default and html modes use pf.full_summary + pf.footer + pf.all_pages().
Breaking change in v0.4: the free functions parse_parquet_file(), get_summary(), get_pages(), and find_footer_segment() were removed. Use the ParquetFile methods above as drop-in replacements.
segment_to_json, json_encode, and fill_gaps remain as module-level utilities.
Page-level decoders for the encoded streams a Parquet inspector actually needs to read:
from parquet_analyzer.decoders import (
decompress,
decode_rle_bitpacked_hybrid,
decode_rle_bitpacked_hybrid_stream,
decode_levels,
decode_v1_level_block,
decode_plain,
DecodeStats,
RleBitPackedStream,
RleRun,
BitPackedRun,
)| Function | What it does |
|---|---|
decompress(data, codec, uncompressed_size) |
Decompress a Parquet compressed byte slice. Supports UNCOMPRESSED, SNAPPY, GZIP, ZSTD, LZ4 (legacy Hadoop framed), LZ4_RAW. |
decode_rle_bitpacked_hybrid_stream(data, bit_width, num_values) |
Decode a raw RLE / bit-packed-hybrid stream into an RleBitPackedStream — the encoding's own structure: bit_width plus the ordered runs (each an RleRun(value, length) or BitPackedRun(length, values)) plus the flattened values. This is the faithful "encoder-logical" view used for both level streams and dictionary indices. |
decode_rle_bitpacked_hybrid(data, bit_width, num_values) |
The same stream as a (values, DecodeStats) pair — the flattened values plus a flat run summary. Use decode_rle_bitpacked_hybrid_stream when the ordered, per-run structure matters. |
decode_levels(data, max_level, num_values) |
Decode definition or repetition levels to a flat list. Bit width is derived from max_level. Returns [0] * num_values if max_level == 0 (no level block exists on disk for required columns). |
decode_v1_level_block(data, offset, max_level, num_values) |
V1-data-page helper that handles the 4-byte little-endian length prefix in front of each level block. Returns (levels, new_offset). V2 data pages store level byte lengths in the page header instead — for those, slice the bytes yourself and call decode_levels directly. |
decode_plain(data, parquet_type, num_values, type_length=None) |
Decode PLAIN-encoded values. Supported types: BOOLEAN, INT32, INT64, INT96, FLOAT, DOUBLE, BYTE_ARRAY, FIXED_LEN_BYTE_ARRAY (type_length required). INT96 is returned as 12-byte bytes — interpretation is the caller's responsibility. |
End-to-end example — decode the indices of a dictionary-encoded V1 data page
and recover the original values. For most callers, Page.decode() (above)
does exactly this; reach for the raw decoders only when you need a custom
decode path. The byte ranges come from the public Page API:
from parquet_analyzer import ParquetFile
from parquet_analyzer.decoders import (
decompress, decode_v1_level_block, decode_rle_bitpacked_hybrid, decode_plain,
)
pf = ParquetFile("example.parquet")
cc = pf.row_groups[0].columns[0] # a dict-encoded BYTE_ARRAY column
dict_page = next(p for p in cc.pages() if p._kind == "dictionary_page")
data_page = cc.page(cc.num_pages - 1)
# Dictionary entries are PLAIN values of the column's physical type.
dict_raw = decompress(dict_page.raw_body(), cc.codec, dict_page.uncompressed_size)
dict_values = decode_plain(dict_raw, cc.type, dict_page.num_values)
# V1 data page layout: [4-byte LE def-level block length][def-level RLE stream]
# [1-byte indices bit-width][indices RLE/bit-packed stream]
raw = decompress(data_page.raw_body(), cc.codec, data_page.uncompressed_size)
def_levels, after_def = decode_v1_level_block(
raw, 0, cc.max_definition_level, data_page.num_values
)
indices_bit_width = raw[after_def]
# In dict-encoded pages, the indices stream contains entries only for
# non-null rows (nulls are represented in the def-level stream above and
# carry no value). So the index count is `num_non_null`, NOT `num_values`.
num_non_null = sum(1 for d in def_levels if d == cc.max_definition_level)
indices, stats = decode_rle_bitpacked_hybrid(
raw[after_def + 1:], indices_bit_width, num_non_null,
)
print(f"RLE runs: {stats.rle_run_count}, bit-packed runs: {stats.bit_packed_run_count}")
# Reassemble nulls into the final value list using the def-level stream.
it = iter(indices)
values = [
dict_values[next(it)] if d == cc.max_definition_level else None
for d in def_levels
]Gotchas worth knowing:
- V1 vs V2 level layout. V1 prefixes each level block with a 4-byte LE length. V2 stores the byte length in the page header instead (see the next bullet for V2's body layout). Use
decode_v1_level_blockfor V1; for V2 slice the bytes yourself and calldecode_levels. - V2 page body layout is
[rep_levels][def_levels][values]. Levels are stored uncompressed regardless ofis_compressed; only the values section is (optionally) compressed. Concretely, V2 callers should sliceraw[:rep_len],raw[rep_len:rep_len + def_len], andraw[rep_len + def_len:]usingrepetition_levels_byte_lengthanddefinition_levels_byte_lengthfrom the page header, then calldecode_levelsdirectly on the level slices (anddecompresson the values slice only whenis_compressedis true). Do not pass an entire V2 page body todecompressas one block. - Required columns have no level block. When
max_def_level == 0, the file contains no def-level bytes at all.decode_levelsreturns[0] * num_valuesin that case. - Indices skip nulls. For nullable columns, the indices stream length is the non-null row count (
sum(def_levels)whenmax_def_level == 1), not the page'snum_values. Askingdecode_rle_bitpacked_hybridfor too many values raisestruncated(or silently produces garbage indices if the indices block is followed by other bytes). - Per-page dictionary-index bit width. Dictionary-encoded data pages start their indices block with a 1-byte
bit_widthchosen by the writer (it may exceedceil(log2(dict_size))). Read that byte first, then pass the rest todecode_rle_bitpacked_hybrid.
The tool uses a custom Thrift protocol implementation (OffsetRecordingProtocol) that wraps the standard Thrift compact protocol to track byte offsets and lengths of all decoded structures. This enables precise mapping of logical Parquet structures to their binary representation.
Because each command is a one-shot process, a complex footer (hundreds of row groups × tens of columns) would re-pay the offset-recording decode — seconds — on every invocation. To keep repeated inspection of the same file fast, the parsed footer of a large file is cached on disk and served on the next open of an unchanged file (e.g. an interactive sequence of commands drops from seconds to tens of milliseconds per call).
The cache is transparent and safe by construction:
- Content-addressed on the footer bytes, so it is self-invalidating — any change to the file produces a different key and a stale entry can never be served. The cache also invalidates automatically across tool/Thrift upgrades.
- Stored in a private (
0700) per-user cache directory; if that directory cannot be confirmed private, the cache silently disables itself and parsing proceeds normally. - Written only for footers large enough to be worth it, and bounded in total size (oldest entries evicted).
It is fully optional:
| Environment variable | Effect |
|---|---|
PARQUET_ANALYZER_NO_CACHE=1 |
Disable the cache entirely (always parse). |
PARQUET_ANALYZER_CACHE_DIR=<path> |
Relocate the cache directory (default: $XDG_CACHE_HOME/parquet-analyzer/ or ~/.cache/parquet-analyzer/). |
PARQUET_ANALYZER_CACHE_MAX_BYTES=<n> |
Eviction ceiling for the cache directory (default 1 GiB). |
The ParquetFile(path, use_cache=False) keyword bypasses it programmatically.
pip install -e .[dev]
hatch run dev:check # will format, lint, type-check, test with coverageThe development extra pulls in tooling (hatch, ruff, pytest, pytest-benchmark) and pyarrow / numpy so tests can generate Parquet fixtures on the fly.
Benchmark tests live under tests/bench/ and are excluded from the default pytest run (and from hatch run dev:check) — they're heavier than the unit tests and not meant for every commit. Run them explicitly:
# Run all benchmarks (synthetic parquet fixtures generated on the fly)
pytest tests/bench/ --benchmark-only
# Compare against the committed baseline (captured at the start of the
# lazy-core work; see tests/bench/baselines/ for the JSON).
# Note: the leading "0001_" is pytest-benchmark's auto-prepended save-
# sequence prefix; --benchmark-compare matches by prefix, so the full
# filename stem is required.
# The -W flag suppresses pytest-benchmark's "machine_info changed"
# warning, which fires noisily on virtualized hosts where cpu.hz_actual
# jitters between runs even on the same machine.
pytest tests/bench/ --benchmark-only \
--benchmark-storage=file://tests/bench/baselines \
--benchmark-compare=0001_eager-v0.4.0 \
-W ignore::pytest_benchmark.logger.PytestBenchmarkWarningThe synthetic fixture generator (tests/bench/generate.py) produces three shapes that stress different lazy-parsing boundaries:
wide— many columns, few rows (footer is large relative to body)tall— few columns, many rows (body is huge relative to footer; the canonical case for the lazy-core's footer-only fast path)deep— few columns, many rows split across many row groups (stresses per-row-group / per-chunk walking)
To capture a fresh baseline (e.g., after a behavior change in the eager path):
pytest tests/bench/ --benchmark-only \
--benchmark-storage=file://tests/bench/baselines \
--benchmark-save=eager-vX.Y.ZBaseline files are platform-specific (CPU, Python version) — pytest-benchmark stores them under tests/bench/baselines/<platform>/.
⚠️ Cross-machine comparisons are not meaningful. The committed baseline (eager-v0.4.0) was captured on the maintainer's machine (Linux/CPython 3.14/x86_64, AMD EPYC 9V74). The per-platform subdirectory prevents the most obvious mismatches (Linux vs macOS, x86 vs ARM), but does not distinguish between CPU SKUs in the same broad category — an old i7 and a current EPYC both land inLinux-CPython-3.14-64bit/, with wildly different absolute numbers. When validating perf changes against the lazy-core work, capture your own baseline before the change and compare against that, not against the committed one. The committed baseline is useful for the maintainer's own session-to-session comparisons and as a frozen reference for the perf claims in PR #18. There's no CI gating on benchmark numbers; perf validation is operator-driven on a single machine at a time.
The Python modules in src/parquet are generated from parquet.thrift.
-
Install the Apache Thrift compiler (
brew install thrifton macOS, or download a release from the Apache Thrift project). -
From the repository root, regenerate everything in one step:
hatch run dev:update-thrift
This refreshes
parquet.thrift, runs the compiler, and removes any straysrc/__init__.pythe compiler may create.
Contributions are welcome! Please feel free to submit issues, feature requests, or pull requests.
This project is licensed under the Apache License 2.0.
© 2025 Chungmin Lee
{ "$schema": "parquet-analyzer/v1/show", "_kind": "row_group", "_location": {"offset": 620, "length": 210}, "num_rows": 10, "columns": [ {"_kind": "column_chunk", "_location": {"offset": 622, "length": 112}, "_path": "row_groups/0/columns/0", "name": "id"} ], "_navigation": {"path": "row_groups/0", "parent": "", "kind": "row_group"} }