0 → N machines
Batch probe queue
Now → machines
InS3 file list, video indexes
WorkProbe format / duration; split shards and jobs
OutShard ranges → register media jobs
CLOUD VIDEO PIPELINE / ELASTIC COMPUTE
The platform stays up. Processing pools scale with the work.
CPU planning, GPU media processing, and CPU containers calling the VLM API — each pool scales on its own.
Always-on control plane · ECS / Fargate Pools can scale to 0. The platform stays online.
AWS Batch adds and removes EC2 instances from queued job demandPlatform scaler → ECS adds and removes containers from backlog and API quota
0 → N machines
Batch media queue
Now → machines
InNeeded byte ranges from S3 source video
WorkDecode → CV / proxy encode → upload
OutS3 media artifacts, CV facts, gate decisions
0 → N containers
SQS model-task queue
Now → containers
InSelected video windows, prompts
OutActivity / events / counts → validated commit
Queues carry task IDs only; workers read and write video on S3.CPU / GPU are EC2 instances; the API pool is Fargate containers.
Close-out check→
Ready event→
In progressAdded on demandReclaimed after finishCounts and animation are illustrative; instance counts need benchmarks.
Click a pool or component to see inputs, outputs, and duties
Scale-up waits for machines to start. Scale-down waits for in-flight work to finish. The number after “→” is demand. Batch schedules from actual job resources.
Zoom into one GPU machine: CPU I/O, GPU compute
On one machine, read, compute, and upload run at the same time.
Same-machine CPU · input thread
Read and demux
Read the needed bytes from S3 by time range, demux compressed packets to the decoder, without downloading the whole file.
Input cache (illustrative)Feed B / prefetch the start of C
32 GiB input cache · illustrative
→
Same-machine GPU · media and CV
Decode, recognize, transcode
The GPU decodes frames once; the same frames feed CV and proxy encode.
Hardware decode · NVDECpackets → frames
Scale / sampleframes → tensors
CV inference→ predictions
Proxy encode · NVENC→ encoded packets
→
Same-machine CPU · output thread
Mux and upload
Merge CV results and pick model windows, mux proxy, upload to S3, and hand the artifact list to result acceptance.
Output cache (illustrative)Uploading continuously
8 GiB output cache · illustrative
Normal: the reader prefers feeding B, then prefetches the start of C; GPU processes B; the uploader sends A’s artifacts.
Shard A · upload / commitShard B · computeShard C · prefetch startEach committed shard can be reused
Memory boundOnly buffer needed compressed bytes and decoded frames; never load the whole file.
Recovery boundResume from checkpoints after interruption; skip committed shards; recompute only uncommitted ranges.
Resource boundStart benchmarks at 4 concurrent streams; watch NVDEC / NVENC / GPU / I/O separately.
When data grows, you scale processing capacity
Default per project: 50 helmets × 8 hours. This is an arithmetic scenario, not a completion guarantee.
Media nodes · theoretical ceil
Platform admits budget → Batch requests capacity
Effective node-hours
Estimated from effective throughput
Cluster raw read floor
GB/s
VLM request count
Estimated from selected duration
Model API RPM needed
TPM quota still required
Steady concurrency at 20 s average latency
Theoretical steady-state
Media nodes can scale to 0; the always-on services stay. Estimate media and VLM separately, and leave headroom for close-out aggregation and the report.