content delivery engineers

Abstract: In the hotplug coded caching model, $K$ users fill their caches of size $M$ as a function of $N$ files, but only an unknown subset of $A\leq K$ users is active in the delivery phase. In all regimes where the optimal hotplug tradeoff is known, it depends only on $A$, not on $K$; whether this holds in general is open. We show that: (i) Under uncoded placement, YMA delivery with MAN placement, where offline users are assigned the demands of active users, is optimal. Hence, offline users cause a strict load penalty at intermediate memory sizes, while the MAN placement is optimal for every number of active users; (ii) Under row-wise placement (RWP), where each cache is a function of the user's own row of each file, every classical scheme can be embedded in a Reed-Solomon code. The resulting hotplug scheme has the tradeoff of the classical model with $A$ users, for every $A\leq K$: there is no load penalty, but the placement depends on $K$. We then derive a converse for RWP, which shows that: (a) for $2\leq N\leq A$, the scheme of Gomez-Vilardebo (GV) is RWP optimal for $AM\leq N/(N-1)$, which settles the RWP tradeoff for $N=2$; (b) YMA is RWP optimal at $AM=N$; and (c) for $N=A=3$, GV is RWP optimal for $M\leq 1$. Hence, the Reed-Solomon embeddings of these schemes are optimal among embedded RWP schemes.

Abstract: IP geolocation supports many Internet services and operations, including but not limited to content delivery, fraud detection, content localization, and regulatory compliance. However, assessing the reliability and accuracy of IP geolocation providers is challenging and only partially explored. In this paper, we present a longitudinal study evaluating the consistency between RIPE Atlas latency measurements and location estimates from four popular geoproviders as a proxy for geolocation quality. Analyzing six months of data we observe mostly stable consistency levels through time, with fluctuations largely related to changes to the mix of targets being measured by RIPE Atlas each day. We also observe that consistency levels can differ greatly between geoproviders, with IPinfo leading DB-IP (the next best) by 5.4 percentage points for IPv4 and 11.5 percentage points for IPv6. Finally, the analysis reveals distinct degradation patterns across providers for different subsets of IP addresses. For example, for IPv4 router addresses, MaxMind consistency is 10.7 percentage points worse than in the general case while IPinfo is only 1.6 percentage points worse; for IPv6, the pattern flips and routers are geolocated more consistently than non-routers by all providers.

Abstract: AI-generated content is becoming a rapidly growing class of digital artifacts. Because these artifacts accumulate over time, their exponential growth creates a substantial storage, energy, and infrastructure cost for operators and society. At the same time, GPU compute cost continues to fall rapidly with each hardware generation. This divergence raises a fundamental question: when does on-demand regeneration become cheaper than persistent storage? This paper develops a cost model for comparing persistent storage and on-demand regeneration for AI-generated artifacts. The model accounts for corpus growth, HDD and tape price trends, drive replacement, electricity, request skew, caching, generator FLOPs, and future GPU price-performance improvements. For image generation, our analysis shows that prompt-based regeneration does not become cheaper than storage until around 2040, because every cache miss must still rerun the full prompt-to-artifact generation pipeline. We observe that widely used diffusion-based generation models operate in latent space, creating an alternative point in the cost tradeoff: instead of storing the final artifact or only the prompt, operators can store a compact intermediate representation (IR) and perform cheap on-demand decoding. Our analysis shows that caching combined with IR-based regeneration substantially reduces both storage and compute cost, making it at least 2x cheaper than both full-object storage and prompt-based regeneration even today. On a production image trace with 2.07 billion requests, the same conclusion holds: prompt-based regeneration is over 100x more expensive than storage, while IR-based regeneration reduces total cost to roughly half that of full-object storage while preserving interactive miss latency.