lazynwb 1.0.0.dev5

An attempt to speed-up access to large NWB (Neurodata Without Borders) files stored in the cloud.

lazynwb

Efficient read-only access to tables, time series and metadata across multiple local or cloud-hosted NWB files simultaneously, without loading entire files into memory.

pip install lazynwb

---

Why lazynwb

Work with a project's worth of NWB files

Seamlessly read and concatenate tables across sessions:

import lazynwb

# read the units table from every session into a single DataFrame
df = lazynwb.get_df(
    ['session_1.nwb', 'session_2.nwb', 'session_3.nwb'],
    '/intervals/trials',
)
# each row keeps its source file in the `_nwb_path` column

Efficient dataframe access with projection and predicate pushdown

NWB tables like /units mix single-value metric columns with large array columns (spike_times, waveform_mean). With pynwb, accessing a dataframe means loading everything.

lazynwb provides a Polars plugin that returns a LazyFrame backed by the NWB file. Only the columns and rows you actually use are loaded:

import lazynwb
import polars as pl

lf = lazynwb.scan_nwb('s3://bucket/session.nwb', '/units')

df = (
    lf
    .filter(
        pl.col('presence_ratio') >= 0.95,    # predicate pushdown: skip non-matching rows
        pl.col('location') == 'VISp',
    )
    .select('unit_id', 'spike_times')        # projection pushdown: only fetch these columns
    .collect()
)

For queries that don't need all columns on data that's stored in the cloud, lazynwb can turn an operations that takes minutes into one that takes seconds.

For more details, see the lazy API guide in the Polars documentation.

A simple, consistent API

One interface for local files, S3/GCS/Azure, HDF5 and Zarr: no extra imports or backend-specific code:

import lazynwb

# local HDF5
lazynwb.get_df('my_file.nwb', '/trials')

# remote Zarr
lazynwb.get_df('s3://bucket/session.nwb', '/trials')

# DANDI archive
dandi_sources = lazynwb.get_dandi_sources('000363', version='0.231012.2129')
lf = lazynwb.scan_nwb(dandi_sources, '/trials')

Basic benchmarks

Streaming a single NWB file over HTTPS (Steinmetz 2019, 312 MB HDF5) with a laptop on a typical home internet connection:

Tables: reading /intervals/trials (214 rows, no array columns):

Method	Time
pynwb .to_dataframe()	8.3 s
lazynwb.get_df	5.1 s
lazynwb.scan_nwb	4.1 s

For a table with no array columns (relatively quick to load), all approaches read the same volume of data.

Tables: reading /units (1085 rows, includes large spike_times and waveform_mean arrays):

Method	Time	What it reads
pynwb .to_dataframe()	231 s	all columns (no choice)
lazynwb.get_df(..., exclude_array_columns=False)	282 s	all columns (equivalent)
lazynwb.get_df(..., exclude_array_columns=True)	6 s	scalar columns only
lazynwb.scan_nwb (filter + select)	10 s	filter on scalar columns, then fetch spike_times

When reading all columns, lazynwb and pynwb take roughly the same time: if you need all data in memory, there's no reason to use lazynwb here. The difference is that pynwb always reads everything, while lazynwb lets you choose.

TimeSeries: lick_times (3190 samples, full download):

Method	Time
pynwb	7.3 s
lazynwb.get_timeseries	6.1 s
lazynwb.get_timeseries (metadata only)	6.2 s

Both download the same data, and pynwb also supports lazy access to time series data. The only advantage here is the consistent API.

See benchmarks/streaming_benchmark.py to reproduce or run against your own files:

python benchmarks/streaming_benchmark.py [NWB_PATH]

Schema inference latency for the remote dynamic-routing workload can be measured with benchmarks/schema_benchmark.py:

uv run python benchmarks/schema_benchmark.py
LAZYNWB_SCHEMA_BENCH_JSON=metrics.json uv run python benchmarks/schema_benchmark.py
LAZYNWB_SCHEMA_BENCH_UNITS_SOURCES_FILE=tests/paths.txt uv run python benchmarks/schema_benchmark.py

The schema benchmark uses an isolated temporary catalog cache, defaults to anonymous public object-store access, and reports cold/warm totals plus per-source range GET counts, fetched bytes, and timings. Useful environment variables are documented in the script docstring.

Why not to use lazynwb

• some convenience features of pynwb will not be available, for example object references in tables
• incomplete coverage of the NWB spec. Focussed on the core metadata, TimeSeries and DynamicTable, and tested primarily on ecephys files. Please file an issue if you need support for a particular container.
• you need to write NWB files

---

Quick start

import lazynwb

# read the trials table as a pandas DataFrame
df = lazynwb.get_df('my_file.nwb', '/intervals/trials')

Use get_internal_paths to find available paths if you're not sure what's in a file:

lazynwb.get_internal_paths('my_file.nwb')
# ['/intervals/trials', '/processing/behavior/running_speed', '/units', ...]

Use get when you want lazynwb to choose the return type from the NWB container:

df = lazynwb.get('my_file.nwb', '/units')  # pandas DataFrame by default
ts = lazynwb.get('my_file.nwb', '/processing/behavior/running_speed', exact_path=True)

Force DataFrame materialization, even for a TimeSeries container:

df = lazynwb.get('my_file.nwb', '/processing/behavior/running_speed', exact_path=True, as_df=True)

---

Reading tables

As a pandas or polars DataFrame (get_df)

Returns a pandas DataFrame by default:

df = lazynwb.get_df('my_file.nwb', '/units')

Return a polars DataFrame instead:

df = lazynwb.get_df('my_file.nwb', '/units', as_polars=True)

Select specific columns:

df = lazynwb.get_df('my_file.nwb', '/units', include_column_names=['unit_id', 'location'])

Exclude specific columns:

df = lazynwb.get_df('my_file.nwb', '/units', exclude_column_names=['waveform_mean'])

Large array columns like spike_times and waveform_mean are excluded by default (exclude_array_columns=True). Include them explicitly:

df = lazynwb.get_df('my_file.nwb', '/units', exclude_array_columns=False)

Read a table across multiple files into a single DataFrame:

df = lazynwb.get_df(
    ['file_1.nwb', 'file_2.nwb', 'file_3.nwb'],
    '/intervals/trials',
)

Each row gets _nwb_path, _table_path and _table_index columns to identify its source file and original row index.

As a Polars LazyFrame (scan_nwb)

scan_nwb returns a polars.LazyFrame that reads data on demand. Only the columns and rows you actually use are fetched from disk or the network, which makes it useful for large files or files on cloud storage.

import lazynwb
import polars as pl

lf = lazynwb.scan_nwb('my_file.nwb', '/units')

# filter rows and select columns - only the needed data is read
df = (
    lf
    .filter(pl.col('presence_ratio') >= 0.9)
    .select('unit_id', 'location', 'spike_times')
    .collect()
)

Read across multiple files:

lf = lazynwb.scan_nwb(
    ['file_1.nwb', 'file_2.nwb'],
    '/units',
)
df = (
    lf
    .filter(
        pl.col('amplitude_cutoff') <= 0.1,
        pl.col('isi_violations_ratio') <= 0.5,
    )
    .select('unit_id', 'location', 'spike_times', '_nwb_path')
    .collect()
)

Control schema inference when files have slightly different column types:

lf = lazynwb.scan_nwb(
    nwb_paths,
    '/units',
    infer_schema_length=5,               # only read first 5 files for schema
    schema_overrides={'unit_id': pl.Int64},  # force a column type
)

There's also read_nwb, which is the same as scan_nwb(...).collect():

df = lazynwb.read_nwb(nwb_paths, '/units')  # returns pl.DataFrame

Note: pl.DataFrame has a .to_pandas() method.

Using LazyNWB (PyNWB-like interface)

Access tables and metadata from a single file with familiar attribute names:

nwb = lazynwb.LazyNWB('my_file.nwb')

# tables (returned as pandas DataFrames)
nwb.trials
nwb.units
nwb.epochs
nwb.electrodes

# metadata
nwb.session_id
nwb.session_start_time
nwb.session_description
nwb.identifier
nwb.experiment_description
nwb.experimenter
nwb.lab
nwb.institution
nwb.keywords

Subject metadata:

nwb.subject.age
nwb.subject.sex
nwb.subject.species
nwb.subject.genotype
nwb.subject.subject_id
nwb.subject.strain
nwb.subject.date_of_birth

Get a table as polars:

df = nwb.get_df('/units', as_polars=True)

Use the general accessor when you want tables as DataFrames and TimeSeries as TimeSeries objects:

result = nwb.get('/processing/behavior/running_speed', exact_path=True)
df = nwb.get('/processing/behavior/running_speed', exact_path=True, as_df=True)

Get a summary of everything in the file:

nwb.describe()
# {'identifier': '...', 'session_id': '...', ..., 'paths': ['/acquisition/...', '/units', ...]}

---

Time series

Get a single time series by searching for a name:

ts = lazynwb.get_timeseries('my_file.nwb', search_term='running_speed')

ts.data          # h5py.Dataset or zarr.Array (lazy - not loaded until sliced)
ts.timestamps    # h5py.Dataset or zarr.Array
ts.unit          # e.g. 'cm/s'
ts.rate          # sampling rate, if available
ts.description

Get a time series by exact internal path:

ts = lazynwb.get_timeseries('my_file.nwb', exact_path=True, search_term='/acquisition/lick_sensor_events')

Get all time series in the file:

all_ts = lazynwb.get_timeseries('my_file.nwb', match_all=True)
# dict: {'/acquisition/lick_sensor_events': TimeSeries(...), '/processing/behavior/running_speed': TimeSeries(...), ...}

Also available on a LazyNWB object:

nwb = lazynwb.LazyNWB('my_file.nwb')
ts = nwb.get_timeseries('running_speed')

---

Metadata across files

Get session and subject metadata for many files at once:

df = lazynwb.get_metadata_df(nwb_paths)  # pandas DataFrame

df = lazynwb.get_metadata_df(nwb_paths, as_polars=True)  # polars DataFrame

Returns columns including identifier, session_id, session_start_time, session_description, subject_id, age, sex, species, genotype, strain, date_of_birth, _nwb_path, and more.

---

File contents and schema

Discover internal paths

See what's inside an NWB file:

paths = lazynwb.get_internal_paths('my_file.nwb')
# ['/acquisition/lick_sensor_events',
#  '/intervals/trials',
#  '/processing/behavior/running_speed',
#  '/units',
#  ...]

path_info = lazynwb.get_internal_path_info('my_file.nwb')
# {'/acquisition/lick_sensor_events': {'is_timeseries': True, 'is_group': True, ...},
#  '/intervals/trials': {'is_group': True, 'attrs': {'colnames': ...}, ...},
#  '/units': {'is_group': True, 'attrs': {'colnames': ...}, ...},
#  ...}

Get table schema

Get the unified column names and types for a table across multiple files:

schema = lazynwb.get_table_schema(nwb_paths, '/intervals/trials')
# OrderedDict([('condition', String), ('id', Int64), ('start_time', Float64), ...])

Uses polars (Arrow) data types.

---

Format conversion

Export NWB tables to other file formats with convert_nwb_tables.

Supported formats: parquet, csv, json, excel, feather, arrow, avro, delta.

output_paths = lazynwb.convert_nwb_tables(
    nwb_paths,
    output_dir='./output',
    output_format='parquet',
)
# {'/intervals/trials': PosixPath('./output/trials.parquet'),
#  '/units': PosixPath('./output/units.parquet')}

Pass format-specific options via keyword arguments:

# parquet with zstd compression
lazynwb.convert_nwb_tables(nwb_paths, './output', output_format='parquet', compression='zstd')

# csv with custom separator
lazynwb.convert_nwb_tables(nwb_paths, './output', output_format='csv', separator='\t')

# json, pretty-printed
lazynwb.convert_nwb_tables(nwb_paths, './output', output_format='json', pretty=True)

Only export tables present in all files:

lazynwb.convert_nwb_tables(nwb_paths, './output', min_file_count=len(nwb_paths))

Use full internal paths as filenames (e.g. intervals_trials.parquet instead of trials.parquet):

lazynwb.convert_nwb_tables(nwb_paths, './output', full_path=True)

---

SQL queries

Register all tables from NWB files as a Polars SQL context:

ctx = lazynwb.get_sql_context(nwb_paths)
df = ctx.execute("SELECT unit_id, location FROM units WHERE presence_ratio > 0.9").collect()

---

Cloud and remote files

All functions accept S3, GCS, Azure Blob Storage and HTTP/HTTPS paths in addition to local file paths:

# S3
df = lazynwb.get_df('s3://my-bucket/data/file.nwb', '/units')

# Google Cloud Storage
df = lazynwb.get_df('gs://my-bucket/data/file.nwb', '/units')

# Azure Blob Storage
df = lazynwb.get_df('az://my-container/data/file.nwb', '/units')

# HTTP/HTTPS
df = lazynwb.get_df('https://example.com/data/file.nwb', '/units')

Configure global defaults via lazynwb.config:

from lazynwb import config

config.use_polars = True                          # return Polars by default from get_df/get_metadata_df
config.use_obstore = True                         # use obstore for S3/GCS/Azure (default: False)
config.use_remfile = False                        # use remfile for HTTP byte-range requests (default: True)
config.anon = True                                # anonymous access across backends
config.fsspec_storage_options = {"request_payer": True}  # backend-specific extras if needed
config.disable_cache = False                      # disable FileAccessor caching (default: False)

For normal AWS S3 buckets, the region belongs to the bucket, not the caller's current AWS session. The fast HDF5 range reader discovers and caches bucket regions per bucket, so avoid setting a generic {"region": "..."} for workflows that may mix buckets from different AWS regions. Keep explicit region or endpoint settings for S3-compatible storage such as localstack, MinIO, or R2, where they describe that custom service rather than an AWS bucket location.

---

DANDI archive

Use DANDI as a URI discovery step: resolve NWB asset sources with get_dandi_sources, then pass those sources to the regular lazynwb APIs. This keeps DANDI-specific version and asset resolution separate from table, TimeSeries, and metadata reads.

Draft-only dandisets need an explicit draft version. This example uses the DANDI:001637 draft sample:

import lazynwb

dandi_sources = lazynwb.get_dandi_sources('001637', version='draft')

Known DANDI asset IDs can be resolved directly, avoiding asset listing:

dandi_sources = lazynwb.get_dandi_sources(
    '001637',
    version='draft',
    asset_ids=[
        'ca248278-e1b2-4896-ad1c-900e4506cd04',
        '1e37bc82-fd23-4cb5-a253-e794cea932ba',
    ],
)

Published versions can be pinned for reproducible analyses:

published_sources = lazynwb.get_dandi_sources(
    '000363',
    version='0.231012.2129',
)

Use the resolved sources with get_df for eager table reads. Large array columns are excluded by default, so this reads scalar columns unless you opt in to specific arrays:

units = lazynwb.get_df(
    dandi_sources[:2],
    '/units',
    include_column_names=['id', 'firing_rate'],
)

Use the same sources with scan_nwb when you want Polars projections and filters to bound the remote reads before collection:

import polars as pl

lf = lazynwb.scan_nwb(dandi_sources, '/units', infer_schema_length=2)
units = (
    lf
    .filter(pl.col('firing_rate') > 1.0)
    .select('id', 'firing_rate', 'spike_times', '_nwb_path')
    .collect()
)

Use a single resolved source for TimeSeries discovery and bounded reads. The returned data and timestamps objects stay lazy until you slice them:

timeseries_by_path = lazynwb.get_timeseries(dandi_sources[0], match_all=True)
path, ts = next(iter(timeseries_by_path.items()))
preview = ts.data[:1000]

Use the resolved sources for metadata across the same DANDI asset set:

metadata = lazynwb.get_metadata_df(dandi_sources, as_polars=True)

Performance expectations for DANDI workflows are the same as for other remote NWB files:

• get_dandi_sources reads DANDI asset metadata and resolves object-store sources; it does not read NWB table or array payloads.
• get_df avoids full large array reads by default with exclude_array_columns=True. Use include_column_names or exclude_column_names to keep table reads narrow.
• scan_nwb lets Polars push down .select(...) projections and .filter(...) predicates before .collect(), which is the preferred path for large remote tables.
• TimeSeries data and timestamps are returned as backend arrays. Slice bounded ranges such as ts.data[:1000]; avoid ts.data[:] unless you intend to read the full remote array.
• Metadata and internal-path discovery use bounded catalog reads where possible, so they should not require broad raw data traversal on supported remote HDF5 sources.

Limit DANDI source discovery before handing sources to table or TimeSeries APIs:

dandi_sources = lazynwb.get_dandi_sources(
    dandiset_id='000363',
    version='0.231012.2129',
    max_assets=10,
)
lf = lazynwb.scan_nwb(dandi_sources, '/units', infer_schema_length=2)

Prefer the source-first form for new workflows because the same resolved sources can feed get_df, scan_nwb, get_timeseries, and get_metadata_df.

The opt-in DANDI:001637 integration checks exercise the draft sample against tables, TimeSeries, and metadata without enabling those multi-GB remote reads in the default test suite:

uv run pytest tests/test_dandi.py tests/test_dandi_tables.py tests/test_dandi_timeseries.py tests/test_dandi_metadata.py --run-dandi-integration -m "integration and dandi_sample"

---

Internal columns

When reading tables from multiple files, three columns are added automatically:

Column	Description
_nwb_path	Path to the source NWB file
_table_path	Internal path of the table (e.g. /units)
_table_index	Row index in the original table

These are available as constants: lazynwb.NWB_PATH_COLUMN_NAME, lazynwb.TABLE_PATH_COLUMN_NAME, lazynwb.TABLE_INDEX_COLUMN_NAME.

---