Marketing ops · Production

Netflix scales Match Cutting ML pipeline across its entire catalog using Amber media ML infrastructure

The problem

Netflix's media ML practitioners struggled with inconsistent media access, expensive repeated computations across independent pipelines, and bespoke triggering components that were hard to maintain — preventing the Match Cutting pipeline from scaling beyond single titles.

First attempt

The original Match Cutting pipeline lacked input file standardization causing quality issues for cross-title matching, bespoke triggering components caused unnecessary re-computation and inconsistencies, and the quadratic pair computation made scaling to cross-catalog matching computationally intractable.

Workflow diagram · grounded in source
1
New video file trigger
trigger
“we automatically initiate scoring for new videos as soon as the standardized video encodes are ready. Amber handles the computation in the dependency chain recursively.”
2
Standardized catalog preprocessing
integration
“The entire Netflix catalog is pre-processed and stored for reuse in machine learning scenarios. Match Cutting benefits from this standardization as it relies on homogeneity across videos for proper matching.”
3
Shot segmentation and deduplication
ai_action
“extract a representation (aka embedding) of each file using a video encoder (i.e. an algorithm that converts a video to a fixed-size vector) and use that embedding to identify and remove duplicate shots”
4
Per-shot representation computation
ai_action
“We compute another representation per shot, depending on the flavor of match cutting”
5
Pair similarity scoring
ai_action
“We enumerate all pairs and compute a score for each pair of representations”
6
Embedding storage in Amber
integration
“We store embeddings in Amber, which guarantees immutability, versioning, auditing, and various metrics on top of the feature values”
7
High-scoring pairs served to editors
output
“serve high-scoring pairs to video editors in internal applications via Marken”
Reported outcome

The Amber infrastructure standardized media access, eliminated redundant computation through feature memoization, and enabled Match Cutting to scale across the entire Netflix catalog with automatic triggering for new videos; the GPU training cluster throughput increased 3–5 times.

Reported metrics
Training system throughput3–5 times
Compute cost from feature reusesaving on compute costs
Reported stack
AmberJasperMarkenMetaflowRayConductorMesonDagobahTitusIcebergTrinoCassandraElastic SearchSparkMezzFSS3BagginsFSx
Source
https://netflixtechblog.com/scaling-media-machine-learning-at-netflix-f19b400243
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The Amber infrastructure standardized media access, eliminated redundant computation through feature memoization, and enabled Match Cutting to scale across the entire Netflix catalog with automatic triggering for new…

What tools did this team use?

Amber, Jasper, Marken, Metaflow, Ray, Conductor, Meson, Dagobah, Titus, Iceberg.

What results were reported?

Training system throughput: 3–5 times; Compute cost from feature reuse: saving on compute costs (source-reported, not independently verified).

What failed first in this deployment?

The original Match Cutting pipeline lacked input file standardization causing quality issues for cross-title matching, bespoke triggering components caused unnecessary re-computation and inconsistencies, and the quadr…

How is this marketing ops AI workflow structured?

New video file trigger → Standardized catalog preprocessing → Shot segmentation and deduplication → Per-shot representation computation → Pair similarity scoring → Embedding storage in Amber → High-scoring pairs served to editors.

Related marketing ops cases