cyscale

Cython-accelerated SCALE codec library for Substrate-based blockchains (Polkadot, Kusama, Bittensor, etc.).

A drop-in replacement for py-scale-codec — same scalecodec module name, same public API, compiled with Cython for improved throughput.

Installation

pip install cyscale

Performance

Benchmarked on Apple M-series (Python 3.13) against py-scale-codec 1.2.12. All timings are µs per call; speedup = py ÷ cy.

Primitives and small types

Benchmark	py (µs)	cy (µs)	speedup
u8 decode	3.01	1.16	2.60×
u16 decode	2.99	1.25	2.40×
u32 decode	3.11	1.26	2.47×
u64 decode	3.09	1.22	2.53×
u128 decode	2.91	1.22	2.39×
Compact decode	9.59	4.38	2.19×
bool decode	2.93	1.18	2.48×
H256 decode	2.93	1.21	2.41×
AccountId decode (SS58 format 42)	11.45	6.11	1.87×
Str decode	12.89	5.82	2.22×
(u32, u64, bool) decode	21.96	5.59	3.93×
u32 encode	2.34	1.10	2.13×
u64 encode	2.33	1.20	1.93×
Compact encode	8.85	4.42	2.00×
H256 encode	2.47	1.01	2.44×

Large payloads

Benchmark	py (µs)	cy (µs)	speedup
Vec decode (64 elements)	224.20	98.13	2.28×
Vec decode (1,024 elements)	3217.32	1423.46	2.26×
Vec decode (16,384 elements)	50396.95	22435.45	2.25×
Bytes decode (1 KB)	14.67	6.87	2.14×
Bytes decode (64 KB)	64.15	45.05	1.42×
Bytes decode (512 KB)	379.99	300.14	1.27×
Vec decode (5 events, V10)	301.10	135.32	2.23×
MetadataVersioned decode (V10, 85 KB)	64958.83	29839.68	2.18×
MetadataVersioned decode (V13, 219 KB)	143029.99	65651.69	2.18×
MetadataVersioned decode (V14, 300 KB)	390902.34	183644.98	2.13×
Bittensor metadata + portable registry (254 KB)	443089.47	212345.78	2.09×

Primitives and small types see ~2.0–2.6× speedup. Large metadata decoding sees ~2.1–2.3× speedup — the gain compounds across thousands of recursive decode calls. Raw bulk byte operations (Bytes/Vec<u8>) above ~64 KB are dominated by memcpy and see a reduced ~1.3–1.4× speedup.

AccountId with SS58 encoding shows a 1.87× speedup — the SS58 encoding itself (ss58_encode) is pure Python and limits gains in that path.

batch_decode (cyscale-only API)

batch_decode(type_strings, bytes_list) amortises Python dispatch overhead across a list of decodes. The baseline below is a py-scale-codec decode loop, which is the equivalent operation without this API. Note: bt_decode is excluded from this comparison because it does not perform SS58 encoding — including it without that post-processing step would be unfair.

Benchmark	py loop (µs)	cy batch (µs)	speedup
batch_decode AccountId ×10	116.66	42.07	2.77×
batch_decode AccountId ×100	1159.47	415.44	2.79×
batch_decode AccountId ×1,000	11530.59	4112.27	2.80×
Mixed (AccountId/u32/u128) ×100	599.73	147.75	4.06×

To reproduce, run:

# save a py-scale-codec baseline
python benchmarks/bench.py --save-baseline benchmarks/baseline_py.json

# compare against cy-scale-codec
PYTHONPATH=. python benchmarks/bench.py --compare benchmarks/baseline_py.json

Examples of different types

Type	Description	Example SCALE decoding value	SCALE encoded value
bool	Boolean values are encoded using the least significant bit of a single byte.	True	0x01
u16	Basic integers are encoded using a fixed-width little-endian (LE) format.	42	0x2a00
Compact	A "compact" or general integer encoding is sufficient for encoding large integers (up to 2**536) and is more efficient at encoding most values than the fixed-width version.	1	0x04
Vec	A collection of same-typed values is encoded, prefixed with a compact encoding of the number of items, followed by each item's encoding concatenated in turn.	[4, 8, 15, 16, 23, 42]	0x18040008000f00100017002a00
str, Bytes	Strings are Vectors of bytes (Vec<u8>) containing a valid UTF8 sequence.	"Test"	0x1054657374
AccountId	An SS58 formatted representation of an account.	"5GDyPHLVHcQYPTWfygtPYeogQjyZy7J9fsi4brPhgEFq4pcv"	0xb80269ec...
Enum	A fixed number of variants, each mutually exclusive. Encoded as the first byte identifying the index of the variant.	{'Int': 8}	0x002a
Struct	For structures, values are named but that is irrelevant for the encoding (only order matters).	{"votes": [...], "id": 4}	0x04b80269...

License

Apache 2.0 — see LICENSE and NOTICE.