November 2026 Summaries

2 posts from Momento

Filter
Month: Year:
Post Summaries Back to Blog
CMCD v2 adds Event Mode, allowing players to POST playback events and periodic reports directly to operator-selected endpoints rather than relying only on data attached to media requests, improving the timing and detail of playback monitoring while shifting ingestion and processing demands to the operator. At large live events, millions of viewers can generate synchronized bursts, retries can amplify outages, and reports may arrive late or out of order, making it difficult to reconstruct sessions and deliver timely insights. FOX addressed similar demands at Super Bowl LIX with a direct beaconing system that processed a peak of 1.8 million messages per second from 15.5 million devices, handling more than 25 billion beacons with no delivery errors and continuous analytics availability. Its architecture used an ingestion gateway, precomputed counters for live dashboards, and compressed Parquet files in cloud storage for analysis. The article applies these lessons to CMCD v2, recommending an architecture that groups reports by session, validates and enriches data, archives raw records, and uses responsive backpressure and configurable reporting rates to manage load. It also emphasizes designing the system around the decisions operators need to make, such as identifying rebuffering problems, adjusting delivery choices, or changing telemetry settings during an incident.
Nov 20, 2026 2,133 words in the original blog post.
Repeatedly processing long inputs in AI inference consumes GPU capacity because attention-state data is often discarded after each request, while local memory and storage may not be large enough to retain it. Inference systems are therefore adding shared context-memory tiers alongside GPU memory, local DRAM, and local SSDs, allowing attention blocks to be reused across workers and clusters. The passage presents Valkey as a potential option in this space, describing a Large Object module developed by Momento and AWS that stores multi-megabyte blocks across DRAM and NVMe and uses RDMA for transfers into worker host memory; it also integrates with LMCache, vLLM, and SGLang. A shared, durable store can preserve cached state when workers or clusters change and may reduce recomputation costs, with the economic case depending on how often data is reused and the relative costs of storage and GPU time. The authors report preliminary performance comparisons favoring Valkey over local NVMe by a small margin, while arguing that tiered storage and shared access can make longer retention practical.
Nov 19, 2026 963 words in the original blog post.