idhash 0.3.0

Calculate the Unique Numeric Fingerprint for Tabular Data

IDHash

Efficiently create an identifiable hash for a dataset that is;

• Independent of row ordering
• Dependent of column ordering

Quickly identify if two datasets are identical by comparing their hashes, without needing to presort the values.

IDHash is based upon UNFv6, and details around UNF can be found here, where the major differences are that UNF is column-invariant but row-dependent, and IDHash is column-dependent and row-invariant.

Example - Hash Entire Dataset

import pandas as pd
import pyarrow as pa
from idhash import id_hash

x=pd.DataFrame.from_dict({'a': [1,2,3]})
print(id_hash(x))
> 128759287229449989335676029604592902962

Example - Hash Iteratively

import pandas as pd
import pyarrow as pa
from idhash import id_hash, create_hasher
from typing import List

x=pd.DataFrame.from_dict({'a': [1,2,3]})
hasher = create_hasher([str(y) for y in x.columns], x.dtypes)
batches = pa.Table.from_pandas(x).to_batches(max_chunksize=1)
for batch in batches:
    hasher.write_batches([batch], delta="Add")
print(hasher.finalize())
> 128759287229449989335676029604592902962

Iterative hashing has an additional benefit - it's possible to verify a delta between two datasets, i.e.

dataset_a: pd.DataFrame, dataset_b: pd.DataFrame, delta: List[pa.RecordBatch] = load_data()
hasher_a = create_hasher(dataset_a.columns, dataset_a.dtypes)
hasher_a.write_batches(pa.Table.from_pandas(dataset_a), delta="Add")
hasher_b = create_hasher(dataset_b.columns, dataset_b.dtypes)
hasher_b.write_batches(pa.Table.from_pandas(dataset_b), delta="Add")

assert hasher_a.write_batches(delta, delta="Add").finalize() == hasher_b.finalize()

Preprocessing

Each column has specific pre-processing according to the UNF definition. This mostly consists of ensuring that floating point values, datetimes, and timestamps are representable consistently across datasets when taking into account floating point epsilon.

Hash Generation

Each row is taken as a single bytestream, and hashed using Murmurhash128. Murmurhash is a non-cryptographically secure hash function that produces a well distributed hash for each individual value. By adding (wrapping around f64::max) the individual hashed primitives, a final hash can be produced for the final dataset that does not take into account duplicates.

Checking for Equality + Delta

As the hashed rows are added to each other to produce the final value, it is also possible to remove rows against the final hash by producing a row hash in the same manner as was originally performed.

Data Processing

IDHash operates over Apache Arrow RecordBatches and can process with zero-copy over the batches.