Inference Logging Client

A Python SDK for decoding MPLog feature logs from proto, arrow, or parquet format. This client enables you to decode binary-encoded feature data from machine learning inference logging pipelines into Spark DataFrames.

It also reads framed .log files produced by asyncloguploader — directly from a local path, an open file-like object, or a gs:// GCS URI — and decodes every embedded MPLog record in one call.

Overview
Installation
Quick Start
Configuration
Core API Reference
Data Types
Supported Feature Types
Encoding Formats Explained
Exception Handling
Command Line Interface
Advanced Usage Examples
Architecture & Internals
Troubleshooting
Development

Overview

The Inference Logging Client is designed to decode MPLog (Model Proxy Log) feature data that has been encoded for efficient storage and transmission. It supports three encoding formats:

Format	Description	Use Case
Proto	Custom binary encoding with generated flag + sequential features	Default, most compact
Arrow	Arrow IPC format with binary columns	Columnar analytics
Parquet	Parquet format with feature map	Long-term storage

In addition, the SDK reads the framed .log container emitted by asyncloguploader. Each .log file holds many MPLog records — the SDK deframes the container and decodes every embedded record (one row per entity) in a single call. See Reading .log files from GCS or local disk.

Key Features

Multi-format support: Decode Proto, Arrow, and Parquet encoded logs
Automatic format detection: Detects encoding format from metadata byte
Zstd compression support: Automatic decompression of zstd-compressed data
Schema fetching: Retrieves feature schemas from inference API with caching
Spark integration: Returns data as PySpark DataFrames
.log container reader: Local path, file-like, or gs:// URI — deframes and decodes in one call
CLI tool: Command-line interface for quick decoding (auto-routes .log and gs:// inputs)
Thread-safe caching: LRU cache for schemas with thread-safe access

Installation

From PyPI

pip install inference-logging-client

From Source

cd py-sdk/inference_logging_client
pip install -e .

With Development Dependencies

pip install -e ".[dev]"

With GCS Support (for `gs://` .log inputs)

pip install "inference-logging-client[gcs]"
# or from source:
pip install -e ".[gcs]"

Dependencies

Package	Version	Purpose
pyspark	==3.3.0	Spark DataFrame operations (exact pin)
pyarrow	>=5.0.0	Arrow/Parquet format support
zstandard	>=0.15.0	Zstd decompression
google-cloud-storage	>=2.0.0	Optional — only required to read `gs://` URIs

Quick Start

Basic Decoding from Bytes

from pyspark.sql import SparkSession
import inference_logging_client

# Create SparkSession
spark = SparkSession.builder \
    .appName("inference-decode") \
    .getOrCreate()

# Read binary MPLog data
with open("inference_log.bin", "rb") as f:
    data = f.read()

# Decode to Spark DataFrame
df = inference_logging_client.decode_mplog(
    log_data=data,
    model_proxy_id="product-ranking-model",
    version=1,
    spark=spark
)

# View the results
df.show()
#    entity_id  feature_price  feature_category  embedding_vector
# 0  prod_123          29.99                 5  [0.1, 0.2, ...]
# 1  prod_456          49.99                 3  [0.3, 0.4, ...]

# Stop SparkSession when done
spark.stop()

Decoding from a Spark DataFrame

from pyspark.sql import SparkSession
import inference_logging_client

# Create SparkSession
spark = SparkSession.builder \
    .appName("inference-decode") \
    .getOrCreate()

# Read parquet file containing MPLog data
df = spark.read.parquet("inference_logs.parquet")

# Expected columns: features, metadata, mp_config_id, entities, ...
print(df.columns)
# ['prism_ingested_at', 'features', 'metadata', 'mp_config_id', 'entities', ...]

# Decode features from each row
decoded_df = inference_logging_client.decode_mplog_dataframe(df, spark)

decoded_df.show()
#    entity_id  prism_ingested_at  mp_config_id  feature_1  feature_2
# 0  user_123   2024-01-15 10:30   my-model      42         3.14
# 1  user_456   2024-01-15 10:30   my-model      17         2.71

spark.stop()

Reading `.log` Files (local disk or GCS)

asyncloguploader writes framed .log containers holding many MPLog records per file, partitioned by model + date + hour:

gs://gcs-dsci-inferflow-async-logger-prd/
└── async-logger-gcs-flush/<model_config_id>/<YYYY-MM-DD>/<HH>/
    └── <pod>_<YYYY-MM-DD>_<HH-MM-SS>_<seq>.log

Decode one file — local path, file-like object, or gs:// URI:

import inference_logging_client as ilc

# Local
pdf = ilc.decode_log_file_to_pandas("./pod.log")

# GCS  (requires: pip install "inference-logging-client[gcs]")
uri = ("gs://gcs-dsci-inferflow-async-logger-prd/async-logger-gcs-flush/"
       "search-ad-head-prepaid/2026-06-30/18/"
       "search-ad-head-prepaid--prd-inferflow-search-ad-ssd-primary-"
       "54c577fb4d-csbhq_2026-06-30_18-29-03_3.log")
pdf = ilc.decode_log_file_to_pandas(uri)

Fan out across a partition (files share the same schema — fetch it once):

from google.cloud import storage
client = storage.Client()
blobs = list(client.list_blobs(
    "gcs-dsci-inferflow-async-logger-prd",
    prefix="async-logger-gcs-flush/search-ad-head-prepaid/2026-06-30/18/",
))
schema = ilc.get_feature_schema("search-ad-head-prepaid", version=4)
for b in blobs:
    ilc.decode_log_file_to_csv(
        f"gs://{b.bucket.name}/{b.name}",
        f"./out/{b.name.rsplit('/', 1)[-1]}.csv",
        schema=schema,
    )

Or use the CLI — same code path, no boilerplate:

inference-logging-client ./pod.log --output-format csv -o pod.csv
inference-logging-client gs://.../pod.log --output-format analyze

Row schema, output-format table, and full API list live in Reading .log Files.

Configuration

Environment Variables

Variable	Default	Description
`INFERENCE_HOST`	`http://localhost:8082`	Inference service base URL
`INFERENCE_PATH`	`/api/v1/inference/mp-config-registry/get_feature_schema`	Schema fetch API path

Setting Environment Variables

export INFERENCE_HOST="https://inference.prod.example.com"
export INFERENCE_PATH="/api/v1/inference/mp-config-registry/get_feature_schema"

Programmatic Configuration

from pyspark.sql import SparkSession
import inference_logging_client

spark = SparkSession.builder.appName("decode").getOrCreate()

# Pass host directly to functions
df = inference_logging_client.decode_mplog(
    log_data=data,
    model_proxy_id="my-model",
    version=1,
    spark=spark,
    inference_host="https://inference.staging.example.com"
)

Core API Reference

decode_mplog()

Main function to decode MPLog bytes to a Spark DataFrame.

def decode_mplog(
    log_data: bytes,
    model_proxy_id: str,
    version: int,
    spark: SparkSession,
    format_type: Optional[Format] = None,
    inference_host: Optional[str] = None,
    decompress: bool = True,
    schema: Optional[list] = None
) -> pyspark.sql.DataFrame:

Parameters

Parameter	Type	Required	Default	Description
`log_data`	`bytes`	Yes	-	The MPLog bytes (possibly zstd compressed)
`model_proxy_id`	`str`	Yes	-	The model proxy config ID for schema lookup
`version`	`int`	Yes	-	The schema version (0-15)
`spark`	`SparkSession`	Yes	-	The SparkSession to use for creating DataFrames
`format_type`	`Format`	No	`None`	Encoding format. If None, auto-detects from metadata
`inference_host`	`str`	No	`None`	Inference service URL. Falls back to `INFERENCE_HOST` env
`decompress`	`bool`	No	`True`	Whether to attempt zstd decompression
`schema`	`list`	No	`None`	Pre-fetched schema to skip API call

Returns

pyspark.sql.DataFrame with:

First column: entity_id - identifier for each entity
Remaining columns: decoded feature values

Exceptions

Exception	When Raised
`ValueError`	Version out of range (0-15)
`ImportError`	Data is zstd-compressed but `zstandard` not installed
`FormatError`	Unsupported format or parse error
`SchemaFetchError`	Failed to fetch schema from API
`SchemaNotFoundError`	No features in schema response

Example: Basic Usage

from pyspark.sql import SparkSession
import inference_logging_client

spark = SparkSession.builder.appName("decode").getOrCreate()

with open("log.bin", "rb") as f:
    data = f.read()

df = inference_logging_client.decode_mplog(
    log_data=data,
    model_proxy_id="recommendation-model",
    version=2,
    spark=spark
)

print(f"Decoded {df.count()} entities with {len(df.columns) - 1} features")

Example: Explicit Format

from pyspark.sql import SparkSession
from inference_logging_client import decode_mplog, Format

spark = SparkSession.builder.appName("decode").getOrCreate()

df = decode_mplog(
    log_data=arrow_encoded_data,
    model_proxy_id="my-model",
    version=1,
    spark=spark,
    format_type=Format.ARROW  # Skip auto-detection
)

Example: Pre-fetched Schema (Performance Optimization)

from pyspark.sql import SparkSession
from inference_logging_client import decode_mplog, get_feature_schema

spark = SparkSession.builder.appName("decode").getOrCreate()

# Fetch schema once
schema = get_feature_schema("my-model", 1, "https://inference.example.com")

# Decode multiple logs with same schema
for log_bytes in batch_of_logs:
    df = decode_mplog(
        log_data=log_bytes,
        model_proxy_id="my-model",
        version=1,
        spark=spark,
        schema=schema  # Reuse cached schema
    )
    process(df)

decode_mplog_dataframe()

Decode MPLog features from a Spark DataFrame containing encoded feature data.

def decode_mplog_dataframe(
    df: pyspark.sql.DataFrame,
    spark: SparkSession,
    inference_host: Optional[str] = None,
    decompress: bool = True,
    features_column: str = "features",
    metadata_column: str = "metadata",
    mp_config_id_column: str = "mp_config_id"
) -> pyspark.sql.DataFrame:

Parameters

Parameter	Type	Required	Default	Description
`df`	`pyspark.sql.DataFrame`	Yes	-	Input Spark DataFrame with MPLog columns
`spark`	`SparkSession`	Yes	-	The SparkSession to use
`inference_host`	`str`	No	`None`	Inference service URL
`decompress`	`bool`	No	`True`	Attempt zstd decompression
`features_column`	`str`	No	`"features"`	Column containing encoded features
`metadata_column`	`str`	No	`"metadata"`	Column containing metadata byte
`mp_config_id_column`	`str`	No	`"mp_config_id"`	Column containing model proxy ID

Expected Input DataFrame Columns

Column	Type	Required	Description
`features`	`bytes/str`	Yes	Encoded feature bytes (raw, base64, or hex)
`metadata`	`int/bytes`	Yes	Metadata byte for version/format detection
`mp_config_id`	`str`	Yes	Model proxy config ID
`entities`	`list/str`	No	Entity IDs (JSON list or single value)
`prism_ingested_at`	`datetime`	No	Preserved in output
`prism_extracted_at`	`datetime`	No	Preserved in output
`created_at`	`datetime`	No	Preserved in output
`parent_entity`	`str/list`	No	Preserved in output
`tracking_id`	`str`	No	Preserved in output
`user_id`	`str`	No	Preserved in output
`year`, `month`, `day`, `hour`	`int`	No	Partition columns, preserved

Returns

pyspark.sql.DataFrame with:

entity_id: Entity identifier (one row per entity)
Metadata columns (if present in input)
Decoded feature columns

Example: Processing Parquet Logs

from pyspark.sql import SparkSession
import inference_logging_client

spark = SparkSession.builder.appName("decode").getOrCreate()

# Read from data lake
df = spark.read.parquet("s3://bucket/inference-logs/dt=2024-01-15/")

# Decode all rows
decoded = inference_logging_client.decode_mplog_dataframe(df, spark)

# Analyze features
decoded.groupBy('mp_config_id').avg('feature_score').show()

Example: Custom Column Names

from pyspark.sql import SparkSession
import inference_logging_client

spark = SparkSession.builder.appName("decode").getOrCreate()

# Your DataFrame has different column names
df = spark.read.csv("custom_logs.csv", header=True)

decoded = inference_logging_client.decode_mplog_dataframe(
    df,
    spark,
    features_column="encoded_data",      # Custom name
    metadata_column="meta",               # Custom name
    mp_config_id_column="model_id"        # Custom name
)

Reading .log Files

All .log reading routes through one shared iterator (iter_decoded_log_rows), so every output sink emits the same rows — only the container differs. Each function accepts a source that can be a local path, an open binary file-like object, or a gs://bucket/key URI.

Available sinks

Function	Returns / Writes	Best for
`decode_log_file(source, spark, ...)`	`pyspark.sql.DataFrame`	Big files, distributed downstream
`decode_log_file_to_pandas(source, ...)`	`pandas.DataFrame`	Notebooks, mid-sized files
`decode_logs(source, spark, ...)`	`pyspark.sql.DataFrame`	Single file OR a directory / GCS prefix — auto-lists and unions
`decode_logs_to_pandas(source, ...)`	`pandas.DataFrame`	Same, pandas flavour
`list_log_sources(source)`	`list[str]`	Enumerate `.log` files under a dir, `gs://` prefix, glob pattern, or list (any mix of the above, deduped)
`decode_log_file_to_csv(source, path, ...)`	rows written (int)	Handoff to BigQuery / Snowflake / Excel
`decode_log_file_to_jsonl(source, path, ...)`	rows written (int)	Streaming pipelines
`write_parsed_log(source, path, ...)`	records written (int)	Human-readable inspection (asynclogparser `.parsed.log` layout)
`analyze_log_file(source)` + `print_analysis(...)`	dict / prints	Structural diagnosis, no schema fetch
`iter_decoded_log_rows(source, ...)`	`Iterator[dict]`	Custom sinks (Kafka, ClickHouse, ...)
`iter_log_records(source, ...)`	`Iterator[(ts_ns, bytes)]`	Raw MPLog payloads for custom decoding
`read_log_file(source, ...)`	`list[(ts_ns, bytes)]`	Small files, tests

Single file vs whole hour partition

decode_logs / decode_logs_to_pandas / decode_logs_distributed all accept any of these — they figure out what you meant and produce one merged DataFrame:

`source`	Behaviour
Single `.log` file (path or `gs://...log`)	Decodes that file.
Local directory	Non-recursive listing of `*.log` inside, unions results.
`gs://bucket/prefix/` (or any URI not ending in `.log`)	Lists every `.log` under the prefix (works at hour, day, or month level).
Glob — path/URI containing `*`, `?`, or `[]`	Resolved segment-by-segment (shell-glob semantics — `` doesn't cross `/`). Example: `gs://.../search-organic-/2026-07-02/15/`.
`list[…]`	Each entry is recursively expanded (concrete file / dir / prefix / glob), results are unioned and deduplicated. Mix any of the forms above.

import inference_logging_client as ilc

# One hour's worth of logs → one Spark DataFrame
df = ilc.decode_logs(
    "gs://gcs-dsci-inferflow-async-logger-prd/async-logger-gcs-flush/"
    "search-organic-l2-ranker-prepaid-rtp-mall-hasp_scaleup/2026-06-30/23/",
    spark,
)

# Same, pandas
pdf = ilc.decode_logs_to_pandas("./hour_dump/")

# See what would be enumerated without decoding
files = ilc.list_log_sources("gs://bucket/prefix/")   # → list[str]

# Multiple configs at once — union of every listed prefix
df = ilc.decode_logs([
    "gs://.../search-organic-l2-ranker-.../2026-07-02/15/",
    "gs://.../search-ad-head-prepaid/2026-07-02/15/",
    "gs://.../fy-ad_dl_multi_pctr_.../2026-07-02/15/",
], spark)

# Glob at the model level — matches every model_config_id whose name
# fits the shell-style pattern. Runs in seconds, not hours, because
# resolution is segment-by-segment via GCS's delimited listing.
df = ilc.decode_logs_distributed(
    "gs://.../search-organic-*/2026-07-02/15/",
    spark,
)

# Mix concrete paths, prefixes, and globs in one list.
df = ilc.decode_logs_distributed([
    "gs://.../search-ad-*/2026-07-02/15/",     # glob
    "gs://.../specific-model/2026-07-02/",     # concrete prefix
    "gs://.../another-pod_..._4328.log",       # concrete file
], spark)

# Date / hour range filters — inclusive on both ends.
df = ilc.decode_logs_distributed(
    "gs://.../<model>/",
    spark,
    date_from="2026-07-01", date_to="2026-07-03",     # multi-day window
    hour_from=10, hour_to=15,                          # business hours only
)

# Parallelise per-file decode across worker processes
pdf = ilc.decode_logs_to_pandas("gs://.../2026-06-30/23/", max_workers=8)
df  = ilc.decode_logs("gs://.../2026-06-30/23/", spark,   max_workers=8)

How it decodes internally (both Spark and pandas variants):

Every file is decoded to a Python list of row-dicts on the driver (sequential when max_workers <= 1; ProcessPoolExecutor fan-out otherwise — workers do pure CPU-bound Python decoding, no Spark).
All rows are merged and passed to one createDataFrame / pd.DataFrame(...) call at the end. No unionByName chain, no quadratic Catalyst planning on hour-level partitions.

Trade-off: rows accumulate in driver Python memory before the single build. Comfortable up to ~1M rows on a typical Databricks driver; beyond that, prefer a per-file loop that writes each result to parquet incrementally.

Pre-fetch the schema once with get_feature_schema(...) and pass it via schema= so workers don't each hit Horizon-v2 on their first record.

The CLI picks up directories and GCS prefixes automatically — pass one to inference-logging-client and it dispatches through decode_logs.

Common keyword arguments

Every decoding function above accepts these — same defaults, same meaning:

Argument	Default	Meaning
`inference_host`	`INFERENCE_HOST` env, else `http://localhost:8082`	Schema API base URL
`decompress`	`True`	Attempt zstd decompression per MPLog payload
`schema`	`None`	Pre-fetched `list[FeatureInfo]`; skips per-record schema fetch
`needed_columns`	`None`	Subset of feature names to keep
`strict`	`False`	`True` raises `FormatError` on any malformed frame

Row shape (every decoding sink)

entity_id, timestamp_ns, mp_config_id, version, format_type,
user_id, tracking_id, parent_entity, <feature_1>, <feature_2>, ...

One row per (record, entity). The Spark sink stringifies list/bytes values for schema stability; all other sinks preserve native Python types.

Frame layout

Frame header (8B, little-endian):  capacity (uint32) + valid_data_bytes (uint32)
Frame body (capacity − 8 bytes):   first valid_data_bytes carry records, rest is padding
Record:                            [4B length][8B timestamp_ns][MPLog protobuf]

Records are packed back-to-back with no per-record alignment. The MPLog protobuf is the same payload get_mplog_metadata() and decode_mplog() consume — its metadata byte still drives per-record format selection (proto / arrow / parquet).

Examples

import inference_logging_client as ilc

# Spark
from pyspark.sql import SparkSession
spark = SparkSession.builder.getOrCreate()
df = ilc.decode_log_file("gs://bucket/pod.log", spark)

# pandas (no Spark)
pdf = ilc.decode_log_file_to_pandas("./pod.log")

# Write to disk
ilc.decode_log_file_to_csv  ("./pod.log", "./out.csv")
ilc.decode_log_file_to_jsonl("./pod.log", "./out.jsonl")
ilc.write_parsed_log        ("./pod.log", "./out.parsed.log")   # asynclogparser format

# Structural diagnosis (no schema fetch, no network)
ilc.print_analysis(ilc.analyze_log_file("./pod.log"))

# CI fail-fast on corrupt frames
ilc.decode_log_file_to_pandas("./golden.log", strict=True)

# Custom sink via the shared iterator
for row in ilc.iter_decoded_log_rows("gs://bucket/pod.log"):
    producer.send("decoded", row)

Parity with `asynclogparser`

asynclogparser	SDK equivalent
`python asynclogparse.py foo.log`	`write_parsed_log("foo.log", "foo.parsed.log")`
`python asynclogparse.py --analyze foo.log`	`analyze_log_file(...)` + `print_analysis(...)`
`deframe_log_file`	`iter_log_records(...)` — byte-level identical
Per-record protobuf parse	`parse_mplog_protobuf(...)`
Per-entity `decode_proto_features`	Auto-dispatched by `format_type` (proto/arrow/parquet)

Additions the SDK carries over the standalone script: zstd decompression of MPLog payloads, cross-file schema cache keyed by (mp_config_id, version), and strict mode for CI. The ~250 lines of byte-scan recovery heuristics in asynclogparser are intentionally not carried — use strict=True for fail-fast, or the default strict=False to warn-and-skip on a bad frame.

get_mplog_metadata()

Extract metadata from MPLog bytes without full decoding. Useful for inspecting format and version.

def get_mplog_metadata(
    log_data: bytes,
    decompress: bool = True
) -> DecodedMPLog:

Parameters

Parameter	Type	Required	Default	Description
`log_data`	`bytes`	Yes	-	The MPLog bytes
`decompress`	`bool`	No	`True`	Attempt zstd decompression

Returns

DecodedMPLog dataclass with:

user_id: User identifier
tracking_id: Request tracking ID
model_proxy_config_id: Model proxy config ID
entities: List of entity IDs
parent_entity: List of parent entity IDs
metadata_byte: Raw metadata byte
compression_enabled: Whether compression was enabled
version: Schema version (0-15)
format_type: Format type int (0=proto, 1=arrow, 2=parquet)

Example: Inspect Log Before Decoding

import inference_logging_client

with open("unknown_log.bin", "rb") as f:
    data = f.read()

metadata = inference_logging_client.get_mplog_metadata(data)

print(f"Model: {metadata.model_proxy_config_id}")
print(f"Version: {metadata.version}")
print(f"Format: {inference_logging_client.get_format_name(metadata.format_type)}")
print(f"Compression: {'enabled' if metadata.compression_enabled else 'disabled'}")
print(f"Entities: {len(metadata.entities)}")

get_feature_schema()

Fetch feature schema from the inference API with automatic caching.

def get_feature_schema(
    model_config_id: str,
    version: int,
    inference_host: Optional[str] = None,
    api_path: Optional[str] = None
) -> list[FeatureInfo]:

Parameters

Parameter	Type	Required	Default	Description
`model_config_id`	`str`	Yes	-	Model proxy config ID
`version`	`int`	Yes	-	Schema version
`inference_host`	`str`	No	`None`	Inference service URL
`api_path`	`str`	No	`None`	API path for schema endpoint

Returns

list[FeatureInfo]: List of feature definitions with:

name: Feature name
feature_type: Feature data type (e.g., "FP32", "INT64", "FP32VECTOR")
index: Feature index in encoding order

Caching Behavior

Schemas are cached by (model_config_id, version) tuple
Cache is thread-safe (uses threading.Lock)
Maximum 100 cached schemas (LRU eviction)
Host/path are NOT part of cache key (schemas are canonical)

Example: Manual Schema Fetch

from inference_logging_client import get_feature_schema

schema = get_feature_schema(
    model_config_id="product-ranking",
    version=3,
    inference_host="https://inference.example.com"
)

for feature in schema:
    print(f"  {feature.index}: {feature.name} ({feature.feature_type})")

clear_schema_cache()

Clear the internal schema cache. Useful for testing or when schemas have changed.

def clear_schema_cache() -> None:

Example

from inference_logging_client import clear_schema_cache, get_feature_schema

# Clear before testing
clear_schema_cache()

# This will fetch fresh from API
schema = get_feature_schema("my-model", 1)

Data Types

Format Enum

from inference_logging_client import Format

class Format(Enum):
    PROTO = "proto"     # Custom binary encoding
    ARROW = "arrow"     # Arrow IPC format
    PARQUET = "parquet" # Parquet format

FeatureInfo

from inference_logging_client import FeatureInfo

@dataclass
class FeatureInfo:
    name: str           # Feature name (e.g., "user_embedding")
    feature_type: str   # Type string (e.g., "FP32VECTOR")
    index: int          # Position in encoded data

DecodedMPLog

from inference_logging_client import DecodedMPLog

@dataclass
class DecodedMPLog:
    user_id: str = ""
    tracking_id: str = ""
    model_proxy_config_id: str = ""
    entities: list[str] = field(default_factory=list)
    parent_entity: list[str] = field(default_factory=list)
    metadata_byte: int = 0
    compression_enabled: bool = False
    version: int = 0
    format_type: int = 0  # 0=proto, 1=arrow, 2=parquet

Supported Feature Types

Scalar Types

Type Aliases	Size	Description
`INT8`, `I8`	1 byte	Signed 8-bit integer
`INT16`, `I16`, `SHORT`	2 bytes	Signed 16-bit integer
`INT32`, `I32`, `INT`	4 bytes	Signed 32-bit integer
`INT64`, `I64`, `LONG`	8 bytes	Signed 64-bit integer
`UINT8`, `U8`	1 byte	Unsigned 8-bit integer
`UINT16`, `U16`	2 bytes	Unsigned 16-bit integer
`UINT32`, `U32`	4 bytes	Unsigned 32-bit integer
`UINT64`, `U64`	8 bytes	Unsigned 64-bit integer
`FP16`, `FLOAT16`, `F16`	2 bytes	IEEE 754 half-precision float
`FP32`, `FLOAT32`, `F32`, `FLOAT`	4 bytes	IEEE 754 single-precision float
`FP64`, `FLOAT64`, `F64`, `DOUBLE`	8 bytes	IEEE 754 double-precision float
`FP8E5M2`, `FP8E4M3`	1 byte	8-bit floating point (raw byte)
`BOOL`, `BOOLEAN`	1 byte	Boolean value

String Types

Type	Description
`STRING`, `STR`	UTF-8 encoded string
`BYTES`	Binary bytes with 2-byte length prefix

Vector Types

All scalar types have vector variants:

Type Pattern	Description
`{TYPE}VECTOR`	e.g., `FP32VECTOR`, `INT64VECTOR`
`VECTOR_{TYPE}`	e.g., `VECTOR_FP32`, `VECTOR_INT64`
`DATATYPE{TYPE}VECTOR`	e.g., `DATATYPEFP32VECTOR`

Vectors can be encoded as:

Binary: Packed element bytes (most common for feature stores)
JSON: JSON array string (fallback)

Example: Working with Vectors

from pyspark.sql import SparkSession
import inference_logging_client

spark = SparkSession.builder.appName("decode").getOrCreate()

df = inference_logging_client.decode_mplog(data, "model", 1, spark)

# Vector columns contain arrays
df.select("entity_id", "user_embedding").show(truncate=False)

# Access vector elements with Spark SQL functions
from pyspark.sql import functions as F
df.select("entity_id", F.element_at("user_embedding", 1).alias("first_elem")).show()

Encoding Formats Explained

Proto Format

The default and most compact encoding format.

┌─────────────────────────────────────────────────────────────┐
│ Byte 0: Generated Flag (1 = no generated values)           │
├─────────────────────────────────────────────────────────────┤
│ Feature 0: [fixed bytes OR 2-byte size + data]              │
├─────────────────────────────────────────────────────────────┤
│ Feature 1: [fixed bytes OR 2-byte size + data]              │
├─────────────────────────────────────────────────────────────┤
│ ...                                                         │
└─────────────────────────────────────────────────────────────┘

Scalars: Fixed size based on type (e.g., 4 bytes for FP32)
Strings/Vectors: 2-byte little-endian size prefix + data

Arrow Format

Uses Arrow IPC (Inter-Process Communication) format.

┌─────────────────────────────────────────────────────────────┐
│ Arrow IPC Stream                                            │
│ ├── Schema: columns "0", "1", "2", ... (binary type)        │
│ └── RecordBatch                                             │
│     ├── Column "0": [entity0_feature0_bytes, ...]           │
│     ├── Column "1": [entity0_feature1_bytes, ...]           │
│     └── ...                                                 │
└─────────────────────────────────────────────────────────────┘

Column names are feature indices as strings ("0", "1", "2", ...)
Each cell contains raw binary feature bytes
All entities in a single IPC blob

Parquet Format

Uses Parquet columnar format.

┌─────────────────────────────────────────────────────────────┐
│ Parquet File                                                │
│ └── Column "Features": map<int, binary>                     │
│     ├── Row 0: {0: bytes, 1: bytes, ...}                   │
│     ├── Row 1: {0: bytes, 1: bytes, ...}                   │
│     └── ...                                                 │
└─────────────────────────────────────────────────────────────┘

Features column is a map from feature index to binary bytes
Each row represents one entity
Alternative: columnar format with index-named columns (like Arrow)

Metadata Byte Layout

Bit Layout:
┌─────┬─────┬─────┬─────┬─────┬─────┬─────┬─────┐
│  7  │  6  │  5  │  4  │  3  │  2  │  1  │  0  │
├─────┴─────┼─────┴─────┴─────┴─────┼─────┼─────┤
│  Format   │       Version         │ Rsv │Comp │
│  (2 bits) │       (4 bits)        │     │     │
└───────────┴───────────────────────┴─────┴─────┘

Format Type:
  00 = Proto
  01 = Arrow
  10 = Parquet
  11 = Reserved

Version: 0-15 (4 bits)
Compression: 0 = disabled, 1 = enabled (zstd)

Exception Handling

Exception Hierarchy

InferenceLoggingError (base)
├── SchemaFetchError     # API request failed
├── SchemaNotFoundError  # No features in response
├── DecodeError          # Feature decoding failed
├── FormatError          # Invalid format or parse error
└── ProtobufError        # Protobuf parsing failed

Example: Comprehensive Error Handling

from pyspark.sql import SparkSession
from inference_logging_client import (
    decode_mplog,
    InferenceLoggingError,
    SchemaFetchError,
    SchemaNotFoundError,
    FormatError,
)

spark = SparkSession.builder.appName("decode").getOrCreate()

try:
    df = decode_mplog(data, "my-model", 1, spark)
except SchemaFetchError as e:
    print(f"Failed to fetch schema: {e}")
    # Check network, inference service availability
except SchemaNotFoundError as e:
    print(f"Schema not found: {e}")
    # Check model_proxy_id and version
except FormatError as e:
    print(f"Invalid data format: {e}")
    # Check data integrity, correct format
except ValueError as e:
    print(f"Invalid parameter: {e}")
    # Check version range (0-15)
except ImportError as e:
    print(f"Missing dependency: {e}")
    # Install zstandard if data is compressed
except InferenceLoggingError as e:
    print(f"Decode error: {e}")
    # Generic fallback
finally:
    spark.stop()

Command Line Interface

Basic Usage

The CLI accepts three kinds of input and auto-routes them:

Input	Detected by	Path
Raw MPLog bytes file or `-` (stdin)	default	`decode_mplog()` — needs `-m` / `-v`
`*.log` container (any local path)	`.log` extension	`decode_log_file()`
`gs://bucket/key` URI	`gs://` prefix	`decode_log_file()`

# Raw MPLog bytes (the original mode)
inference-logging-client --model-proxy-id my-model --version 1 input.bin

# Asyncloguploader .log file — no -m/-v needed (per-record from metadata)
inference-logging-client ./pod-xyz_17-58-12.log -o decoded.csv

# Same, but straight from GCS
inference-logging-client gs://my-bucket/asynclogs/pod-xyz_17-58-12.log --json

CLI Arguments

Argument	Short	Required	Default	Description
`input`	-	Yes	-	MPLog bytes file, `*.log` path, `gs://...` URI, or `-` for stdin
`--model-proxy-id`	`-m`	Raw-bytes mode only	-	Model proxy config ID (ignored for `.log` / `gs://`)
`--version`	`-v`	Raw-bytes mode only	-	Schema version (ignored for `.log` / `gs://`)
`--format`	`-f`	No	`auto`	Format: `proto`, `arrow`, `parquet`, `auto` (raw-bytes mode only)
`--inference-host`	-	No	env/localhost	Inference service URL
`--hex`	-	No	-	Input is hex-encoded (raw-bytes mode only)
`--base64`	-	No	-	Input is base64-encoded (raw-bytes mode only)
`--no-decompress`	-	No	-	Skip zstd decompression
`--output`	`-o`	No	stdout	Output destination. Spark sink writes a directory (part-files) — the default for `.log`/`gs://` inputs unless `--output-format` says otherwise. Non-Spark sinks (`--output-format csv/jsonl/text/pandas`) write a single file. If `-o foo.csv` / `-o foo.jsonl` / `-o foo.parsed.log` is given without an explicit `--output-format`, the format is auto-inferred from the extension.
`--json`	-	No	-	Output as JSON
`--spark-master`	-	No	`local[*]`	Spark master URL
`--strict`	-	No	-	`.log` mode only: raise on malformed frames instead of skipping
`--output-format`	-	No	`spark` (in .log mode)	`.log`/`gs://` only: `spark`, `pandas`, `csv`, `jsonl`, `text`, or `analyze`

Examples

# Output to CSV directory
inference-logging-client -m my-model -v 1 input.bin -o output_dir

# Output as JSON
inference-logging-client -m my-model -v 1 input.bin --json

# Read from stdin (base64 encoded)
echo "BASE64_DATA" | inference-logging-client -m my-model -v 1 --base64 -

# Read from stdin (hex encoded)
cat hex_data.txt | inference-logging-client -m my-model -v 1 --hex -

# Explicit Arrow format
inference-logging-client -m my-model -v 1 --format arrow input.bin

# Custom inference host
inference-logging-client -m my-model -v 1 \
    --inference-host https://inference.prod.example.com \
    input.bin

# Custom Spark master
inference-logging-client -m my-model -v 1 \
    --spark-master spark://master:7077 \
    input.bin

# Skip decompression (for pre-decompressed data)
inference-logging-client -m my-model -v 1 --no-decompress input.bin

# --- .log / GCS examples -----------------------------------------

# Decode a local asyncloguploader .log file to CSV
inference-logging-client ./pod-xyz_17-58-12.log -o decoded.csv

# Decode straight from GCS (requires the [gcs] extra)
inference-logging-client gs://my-bucket/asynclogs/pod-xyz_17-58-12.log \
    --inference-host https://inference.prod.example.com \
    -o gs_decoded.csv

# CI-style gate: fail on any corrupt frame
inference-logging-client gs://my-bucket/asynclogs/golden.log --strict --json

# Pandas-style CSV print to stdout (no Spark)
inference-logging-client ./pod-xyz_17-58-12.log --output-format pandas

# Write CSV / JSONL directly (no Spark)
inference-logging-client ./pod-xyz_17-58-12.log --output-format csv  -o decoded.csv
inference-logging-client ./pod-xyz_17-58-12.log --output-format jsonl -o decoded.jsonl

# Asynclogparser-compatible human-readable text
inference-logging-client ./pod-xyz_17-58-12.log --output-format text -o decoded.parsed.log

# Structural diagnosis only (no schema fetch, no decoding)
inference-logging-client ./pod-xyz_17-58-12.log --output-format analyze

CLI Output Format

+----------+----------+----------+----------+
| entity_id| feature_1| feature_2| feature_3|
+----------+----------+----------+----------+
| entity_0 |      1.5 |      2.5 |      3.5 |
| entity_1 |      4.5 |      5.5 |      6.5 |
+----------+----------+----------+----------+

--- Summary ---
Format: proto (from metadata)
Version: 1
Compression: disabled
Rows: 2
Columns: 4
Features: feature_1, feature_2, feature_3...

Advanced Usage Examples

Batch Processing with Schema Reuse

import os
import glob
from pyspark.sql import SparkSession
import inference_logging_client

spark = SparkSession.builder.appName("batch-decode").getOrCreate()

# Pre-fetch schema once
schema = inference_logging_client.get_feature_schema(
    "batch-model", 2, "https://inference.example.com"
)

def process_file(filepath):
    with open(filepath, "rb") as f:
        data = f.read()
    
    return inference_logging_client.decode_mplog(
        log_data=data,
        model_proxy_id="batch-model",
        version=2,
        spark=spark,
        schema=schema  # Reuse cached schema
    )

# Process files sequentially
log_files = glob.glob("/data/logs/*.bin")
all_dfs = [process_file(f) for f in log_files]

# Union all DataFrames
from functools import reduce
all_data = reduce(lambda a, b: a.union(b), all_dfs)
print(f"Total entities: {all_data.count()}")

spark.stop()

Feature Analysis Pipeline

from pyspark.sql import SparkSession
from pyspark.sql import functions as F
import inference_logging_client

spark = SparkSession.builder.appName("analysis").getOrCreate()

# Decode logs
df = inference_logging_client.decode_mplog_dataframe(
    spark.read.parquet("logs.parquet"),
    spark
)

# Analyze vector features
embedding_col = "user_embedding"

# Get embedding statistics
df.select(
    F.size(F.col(embedding_col)).alias("dimension"),
    F.aggregate(F.col(embedding_col), F.lit(0.0), lambda acc, x: acc + x).alias("sum")
).show()

# Find entities with unusual embeddings (using array functions)
df.withColumn(
    "embedding_norm",
    F.sqrt(F.aggregate(
        F.col(embedding_col),
        F.lit(0.0),
        lambda acc, x: acc + x * x
    ))
).filter(F.col("embedding_norm") > 10.0).show()

spark.stop()

Integration with Feature Store

from pyspark.sql import SparkSession
import inference_logging_client

spark = SparkSession.builder.appName("feature-compare").getOrCreate()

# Decode inference logs
df = inference_logging_client.decode_mplog(data, "ranking-model", 1, spark)

# Compare with feature store values
from feature_store import FeatureStoreClient

fs = FeatureStoreClient()

# Collect for comparison (for small datasets)
for row in df.collect():
    entity_id = row['entity_id']
    
    # Get fresh features from store
    fresh_features = fs.get_features(entity_id, ["feature_a", "feature_b"])
    
    # Compare logged vs fresh
    for feature_name in ["feature_a", "feature_b"]:
        logged = row[feature_name]
        fresh = fresh_features[feature_name]
        
        if logged != fresh:
            print(f"Drift detected for {entity_id}.{feature_name}:")
            print(f"  Logged: {logged}")
            print(f"  Fresh:  {fresh}")

spark.stop()

Custom Schema Source

from pyspark.sql import SparkSession
from inference_logging_client import decode_mplog, FeatureInfo

spark = SparkSession.builder.appName("custom-schema").getOrCreate()

# Define schema manually (useful for testing or offline processing)
custom_schema = [
    FeatureInfo(name="user_age", feature_type="INT32", index=0),
    FeatureInfo(name="user_score", feature_type="FP32", index=1),
    FeatureInfo(name="user_embedding", feature_type="FP32VECTOR", index=2),
    FeatureInfo(name="user_category", feature_type="STRING", index=3),
]

df = decode_mplog(
    log_data=data,
    model_proxy_id="my-model",  # Not used when schema provided
    version=1,                   # Not used when schema provided
    spark=spark,
    schema=custom_schema
)

spark.stop()

Architecture & Internals

Module Structure

inference_logging_client/
├── __init__.py      # Public API exports, decode_mplog(), decode_mplog_dataframe()
├── __main__.py      # Module execution entry point
├── cli.py           # Command-line interface (auto-routes .log / gs://)
├── decoder.py       # Core byte decoding, type conversion
├── exceptions.py    # Exception classes
├── formats.py       # Proto/Arrow/Parquet format decoders
├── io.py            # Schema fetching, protobuf parsing
├── log_reader.py    # .log frame deframer + GCS reader + decode_log_file()
├── types.py         # Data type definitions (Format, FeatureInfo, DecodedMPLog)
└── utils.py         # Utility functions (type normalization, formatting)

Decoding Flow

                    ┌──────────────────┐
                    │   MPLog Bytes    │
                    │  (compressed?)   │
                    └────────┬─────────┘
                             │
                    ┌────────▼─────────┐
                    │  Zstd Decompress │
                    │   (if enabled)   │
                    └────────┬─────────┘
                             │
                    ┌────────▼─────────┐
                    │ Parse Protobuf   │
                    │ (outer wrapper)  │
                    └────────┬─────────┘
                             │
              ┌──────────────┼──────────────┐
              │              │              │
      ┌───────▼───────┐ ┌────▼────┐ ┌───────▼───────┐
      │  Proto Format │ │  Arrow  │ │ Parquet Format│
      │    Decoder    │ │ Decoder │ │    Decoder    │
      └───────┬───────┘ └────┬────┘ └───────┬───────┘
              │              │              │
              └──────────────┼──────────────┘
                             │
                    ┌────────▼─────────┐
                    │  Feature Schema  │◄──── API Fetch
                    │    (cached)      │
                    └────────┬─────────┘
                             │
                    ┌────────▼─────────┐
                    │  Decode Features │
                    │  (by type)       │
                    └────────┬─────────┘
                             │
                    ┌────────▼─────────┐
                    │ Spark DataFrame  │
                    └──────────────────┘

Schema Cache

# Thread-safe LRU cache with max 100 entries
# Key: (model_config_id, version)
# Value: list[FeatureInfo]

# Cache is NOT keyed by host/path because schemas are canonical
# Same model+version = same schema regardless of which host serves it

Troubleshooting

Common Issues

"No features found in schema"

SchemaNotFoundError: No features found in schema for model_config_id=xxx, version=1

Causes:

Incorrect model_proxy_id
Wrong version number
Schema not yet registered

Solutions:

Verify model_proxy_id matches exactly
Check available versions in inference service
Use get_mplog_metadata() to see the version in the data

"Version out of valid range"

ValueError: Version 16 is out of valid range (0-15)

Cause: Version is encoded in 4 bits (0-15 only)

Solution: Check the version number passed to decode functions

"Data appears to be zstd-compressed but zstandard not installed"

ImportError: Data appears to be zstd-compressed but the 'zstandard' package is not installed.

Solution:

pip install zstandard

"Failed to read Arrow IPC data"

Causes:

Corrupted data
Wrong format specified
Incomplete data

Solutions:

Use format_type=None for auto-detection
Check data integrity
Try get_mplog_metadata() to inspect format

Empty DataFrame Returned

Causes:

No entities in the log
All features decoded as None
Schema mismatch

Solutions:

Check get_mplog_metadata() to verify entity count
Verify schema matches data version
Check for decode warnings

Debug Mode

import warnings
import logging
from pyspark.sql import SparkSession

# Enable all warnings
warnings.simplefilter("always")

# Enable debug logging for HTTP requests
logging.basicConfig(level=logging.DEBUG)

# Create Spark session with verbose logging
spark = SparkSession.builder \
    .appName("debug") \
    .config("spark.driver.extraJavaOptions", "-Dlog4j.logger.org.apache.spark=DEBUG") \
    .getOrCreate()

# Inspect before decoding
import inference_logging_client

metadata = inference_logging_client.get_mplog_metadata(data)
print(f"Format: {metadata.format_type}")
print(f"Version: {metadata.version}")
print(f"Entities: {len(metadata.entities)}")
print(f"Model: {metadata.model_proxy_config_id}")

Development

Setup

# Clone repository
git clone https://github.com/Meesho/BharatMLStack.git
cd BharatMLStack/py-sdk/inference_logging_client

# Create virtual environment
python -m venv venv
source venv/bin/activate

# Install in editable mode with dev dependencies
pip install -e ".[dev]"

Running Tests

pytest

# With coverage
pytest --cov=inference_logging_client --cov-report=html

Code Formatting

# Format with black
black inference_logging_client/

# Lint with ruff
ruff check inference_logging_client/

Building Package

python -m build

License

MIT License

Repository

https://github.com/Meesho/BharatMLStack

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Fork the repository
Create your feature branch (git checkout -b feature/amazing-feature)
Commit your changes (git commit -m 'Add amazing feature')
Push to the branch (git push origin feature/amazing-feature)
Open a Pull Request

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

inference_logging_client-1.0.1.tar.gz (82.4 kB view details)

Uploaded Sep 4, 2026 Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

The dropdown lists show the available interpreters, ABIs, and platforms. Enable javascript to be able to filter the list of wheel files.

inference_logging_client-1.0.1-py3-none-any.whl (57.9 kB view details)

Uploaded Sep 4, 2026 Python 3

File details

Details for the file inference_logging_client-1.0.1.tar.gz.

File metadata

Download URL: inference_logging_client-1.0.1.tar.gz
Upload date: Sep 4, 2026
Size: 82.4 kB
Tags: Source
Uploaded using Trusted Publishing? No
Uploaded via: twine/7.0.0 CPython/3.12.13

File hashes

Hashes for inference_logging_client-1.0.1.tar.gz
Algorithm	Hash digest
SHA256	`bc82523bc13c652a04747c44caa33ae04461634268677d3657b505aaf8cc700a`
MD5	`cfac9b956dd8933c6a5bfac805894696`
BLAKE2b-256	`dfc5b4424c1e99f00394c2842654d994913b35a7bccf094f1ea734d4d4d4eec1`

See more details on using hashes here.

File details

Details for the file inference_logging_client-1.0.1-py3-none-any.whl.

File metadata

Download URL: inference_logging_client-1.0.1-py3-none-any.whl
Upload date: Sep 4, 2026
Size: 57.9 kB
Tags: Python 3
Uploaded using Trusted Publishing? No
Uploaded via: twine/7.0.0 CPython/3.12.13

File hashes

Hashes for inference_logging_client-1.0.1-py3-none-any.whl
Algorithm	Hash digest
SHA256	`d5d1eca5b3ef179c515ac4a87ae05ebb144399b9e00da30fb7f007e7ab13e6de`
MD5	`099968e5b3b41f4e9c5b0589aeca97fa`
BLAKE2b-256	`4e0e6276e9fd6592aabf1cf01b6b9e1572c2408fe28e9aa9f6c4710e1ca90754`