One pass over your data. A self-contained HTML report, a versioned JSON payload, or a CI gate: all from the same pass.
- Run it in your browser →: drop a CSV, Parquet file or Excel workbook and get the real report back. The profiler is compiled to WebAssembly and runs in the page, so nothing is uploaded.
- Open a finished report →: two years of hourly bike rentals, as PySuricata renders it — all four card kinds, and correlations worth ranking.
uv add pysuricata # or: pip install pysuricataimport pandas as pd
from pysuricata import profile
df = pd.read_csv("bike_sharing.csv")
profile(df).save_html("report.html")That is the whole API for the common case. Optional extras:
uv add "pysuricata[polars]" # polars.DataFrame and LazyFrame
uv add "pysuricata[system]" # psutil-backed memory reportingIt reads your data once. Data is processed in chunks using streaming algorithms, so memory usage stays bounded in the number of rows: a million rows costs no more than twenty thousand. It is not bounded in the number of columns, which is a real limit rather than a footnote. See Where the bound stops holding.
It is not only a report. The same pass gives three outputs: profile() for the HTML, summarize() for a versioned JSON payload with no markup in the way, and pysuricata check for a CI gate that exits non-zero when a threshold is crossed. Most profilers give you the first and stop.
Arrow is the boundary, not pandas. Anything exporting the Arrow C stream interface (__arrow_c_stream__) is profiled without materialising it, whatever library produced it. Arrow IPC is what R, Julia and Rust write, so a file from another runtime is read directly.
Approximations say so. Quantiles, distinct counts and duplicate estimates come from sketches. The report labels them and carries their error bound rather than printing an estimate as an exact integer.
One file, no assets. A report is a single HTML file with inline CSS, JS and SVG. It opens from a mail attachment on a machine with no network.
- Streaming architecture: data is processed in configurable chunks, keeping memory bounded in rows, though not in columns (see above). Useful for datasets with more rows than fit in RAM.
- Pandas and Polars: works natively with
pandas.DataFrame,polars.DataFrameandpolars.LazyFrame, plus Parquet files, Arrow IPC files (.arrow,.feather,.ipc), DuckDB relations and Arrow batches. - Configurable: control chunk size, sample size, correlations and more with keyword options, a
preset=, or aProfileConfig. - Reproducible: seeded random sampling produces deterministic results across runs.
- Typed: ships
py.typed;summarize()returns a payload carrying aschema_version. - CLI tool:
profile,summarizeandcheckfrom the command line.
Each column is analyzed based on its type:
- Numeric: mean, variance, skewness, kurtosis, quantiles, histogram, outlier detection (IQR, MAD, z-score), correlations
- Categorical: top values, distinct count, entropy, Gini impurity, string length statistics
- DateTime: temporal range, hour/day/month distributions, monotonicity detection
- Boolean: true/false counts and ratios, entropy
Plus dataset-level metrics: row/column counts, memory usage, missing value percentages, and duplicate row estimates.
The examples below assume a df in scope. The Quick Start frame works, or anything of your own:
import numpy as np
import pandas as pd
rng = np.random.default_rng(0)
df = pd.DataFrame(
{
"age": rng.normal(30, 12, 800).round(1),
"fare": rng.gamma(2, 20, 800).round(2),
"sex": rng.choice(["male", "female"], 800),
"booked": pd.date_range("2024-01-01", periods=800, freq="h"),
}
)Use summarize() for CI/CD quality checks. The payload carries a schema_version and is treated as a contract:
from pysuricata import summarize
stats = summarize(df)
assert stats["schema_version"] == 2
assert stats["dataset"]["missing_cells_pct"] < 5.0
# Gate on the upper bound, not the point estimate. Below the sketch's own
# resolution `duplicate_rows_est` is suppressed to 0 -- a frame with no
# duplicates and one whose duplicates are merely unresolved both read 0,
# and a gate reading either alone would pass the second case by accident.
# `duplicate_rows_hi` is the same bound the HTML report prints either way,
# so this fails closed instead.
rows = stats["dataset"]["rows_est"]
duplicate_pct_hi = stats["dataset"]["duplicate_rows_hi"] / rows * 100 if rows else 0.0
assert duplicate_pct_hi < 1.0
print(f"Mean age: {stats['columns']['age']['mean']:.1f}")Process datasets larger than RAM by passing a generator:
import pandas as pd
from pysuricata import profile
def read_in_chunks():
for i in range(100):
yield pd.read_parquet(f"data/part-{i}.parquet")
report = profile(read_in_chunks())
report.save_html("large_report.html")A Parquet path, an Arrow IPC file, a DuckDB relation or an Arrow source needs no generator at all. Hand it over and it is read a batch at a time, without ever existing as one frame:
import duckdb
from pysuricata import profile
report = profile("data/events.parquet")
# Written by arrow::write_ipc_file() in R, Arrow.write() in Julia, or the
# arrow crate in Rust. The framing is read from the file, not its extension.
report = profile("data/events.arrow")
# A relation is a query that has not run yet, so a filtered join across
# several files is profiled without any of it being landed.
relation = duckdb.connect("warehouse.db").sql("SELECT * FROM events")
report = profile(relation)Measured on a 4,000,000 × 6 frame written as a 180 MB Parquet file, above a 118 MB bare-import floor: 307 MB for profile(path) against 581 MB for profile(pd.read_parquet(path)).
The readers behind that (stream_parquet, stream_ipc, stream_arrow and stream_duckdb) are exported from pysuricata.sources for when you want the batches rather than a profile.
Bounded memory is a claim about rows, not about columns. Each column keeps its own sketches for the whole run and gets its own card in the report, so both memory and report size grow linearly with the width of the frame.
Measured with python -m benchmarks.columns at 20,000 rows, taking the slope from 100
columns to 600: 529 KB of memory and 59 KB of report per column. A 20,000 x 600
frame peaks at 631 MB and emits a 35 MB report, while a 1,000,000 x 14 frame holds 1.2x
the cells and peaks at 344 MB — none of which is the profiling, which never exceeds what
holding the frame already costs. Wide frames are the axis the streaming design does not
yet cover, tracked in #207.
compare() runs both through the same single pass and reports what moved:
from pysuricata import compare
last_week, this_week = df.iloc[:400], df.iloc[400:]
diff = compare(last_week, this_week)
diff.schema.added # columns that appeared
diff.columns["fare"].median_shift_sigma # in baseline standard deviations
diff.to_dict() # JSON-safe, three sectionsIt reports every delta, whether or not it crosses a threshold, because it is a description and not a verdict. pysuricata check is the same arithmetic with a threshold and an exit code.
Pass keyword options for the common cases:
from pysuricata import profile
report = profile(
df,
chunk_size=250_000, # default 50_000
sample=20_000,
seed=42,
correlations=True,
title="My Analysis",
)Or start from a preset, "fast" or "thorough":
from pysuricata import profile
report = profile(df, preset="fast")For everything else, build a ProfileConfig. Keyword options and config= are mutually exclusive:
from pysuricata import profile, ProfileConfig
config = ProfileConfig()
config.compute.chunk_size = 250_000
config.compute.random_seed = 42
config.compute.corr_threshold = 0.5
config.render.title = "My Analysis"
report = profile(df, config=config)See the Configuration Guide for all options.
# Generate an HTML report
pysuricata profile data.csv --output report.html
# Get JSON statistics
pysuricata summarize data.csv
# Compare against a stored baseline; exit non-zero when a threshold is crossed
pysuricata check data.csv --write-baseline baseline.json
pysuricata check data.csv --baseline baseline.json --max-missing-pct 5check exits 0 on pass, 1 when a threshold is crossed, and 2 when the check could not run, so it drops into CI without a wrapper.
The same gate, as a step instead of a script:
- uses: alvarodiez20/pysuricata@v1
with:
file: data.csv
baseline: baseline.json
max-missing-pct: 5Every pysuricata check flag is an input; see action.yml.
PySuricata uses well-known streaming algorithms from the academic literature:
| Algorithm | Purpose | Time | Space |
|---|---|---|---|
| Welford/Pébay | Exact mean, variance, skewness, kurtosis | O(1) per value | O(1) |
| KMV sketch | Distinct count estimation (~2.2% error) | O(log k) per value | O(k) |
| Misra-Gries | Top-k frequent values | O(1) amortized | O(k) |
| Reservoir sampling | Uniform random sample for quantiles | O(1) per value | O(s) |
k = top_k (default 50) for Misra-Gries; max_uniques (default 2048) sizes the KMV distinct-count sketch; s = sample size (numeric_sample_size, default 20 000)
KMV's relative standard error is 1/sqrt(k - 2), which is where the ~2.2% comes from. Approximate values are labelled approximate in the report and carry their error bound rather than being printed as exact integers.
All statistics are computed in a single pass over the data.
- Quick Start
- User Guide
- Configuration
- Arrow, Parquet and DuckDB
- Command Line
- Gating CI on drift
- Comparing two datasets
- API Reference · generated reference
- The
summarize()schema - Statistical Methods
- Examples
Contributions are welcome. See the Contributing Guide.
git clone https://github.com/alvarodiez20/pysuricata.git
cd pysuricata
uv sync --dev
uv run pytestMIT License. See LICENSE for details.
Built using algorithms from:
- Welford, B.P. (1962): streaming moments
- Pébay, P. (2008): parallel merging of moments
- Bar-Yossef, Z. et al. (2002): KMV distinct count estimation
- Misra, J. & Gries, D. (1982): streaming heavy hitters
Named after suricatas (meerkats): small, vigilant animals that work cooperatively and thrive in harsh environments with limited resources.
