Shubham Agarwal

Retrieval-Augmented Generation (RAG) is often used with Large Language Models (LLMs) to infuse domain knowledge or userspecific information. In RAG, given a user query, a retriever extracts chunks of relevant text from a knowledge base. These chunks are sent to an LLM as part of the input prompt. Typically, any given chunk is repeatedly retrieved across user questions. However, currently, for every question, attention-layers in LLMs fully compute the key values (KVs) repeatedly for the input chunks, as state-of-the-art methods cannot reuse KV-caches when chunks appear at arbitrary locations with arbitrary contexts. Naive reuse leads to output quality degradation. This leads to potentially redundant computations on expensive GPUs and increases latency. In this work, we propose Cache-Craft, a system for managing and reusing precomputed KVs corresponding to the text chunks (we call chunk-caches) in RAG-based systems. We present how to identify chunk-caches that are reusable, how to efficiently perform a small fraction of recomputation to fix the cache to maintain output quality, and how to efficiently store and evict chunk-caches in the hardware for maximizing reuse while masking any overheads. With real production workloads as well as synthetic datasets, we show that Cache-Craft reduces redundant computation by 51% over SOTA prefix-caching and 75% over full recomputation. Additionally, with continuous batching on a real production workload, we get a 1.6× speedup in throughput and a 2× reduction in end-to-end response latency over prefix-caching while maintaining quality, for both the LLaMA-3-8B and LLaMA-3-70B models.

Text-to-image diffusion models excel in generating photo-realistic images but are hampered by slow processing times. Training-free retrieval-based acceleration methods, which leverage pre-generated "trajectories," have been introduced to address this. Yet, these methods often lack diversity and fidelity as they depend heavily on similarities to stored prompts. To address this, we present ReCON (Retrieving CONcepts), an innovative retrieval-based diffusion acceleration method that extracts visual "concepts" from prompts, forming a knowledge base that facilitates the creation of adaptable trajectories. Consequently, ReCON surpasses existing retrieval-based methods, producing high-fidelity images and reducing required Neural Function Evaluations (NFEs) by up to 40%. Extensive testing on MS-COCO, Pick-a-pick, and DiffusionDB datasets confirms that RECON consistently outperforms established methods across multiple metrics such as Pick Score, CLIP Score, and Aesthetics Score. A user study further indicates that 76% of images generated by ReCON are rated as the highest fidelity, outperforming two competing methods, a purely text-based retrieval and a noise similarity-based retrieval.

Automated visualization recommendation (Vis-Rec) models help users to derive crucial insights from new datasets. Typically, such automated \visrec models first calculate a large number of statistics from the datasets and then use machine-learning models to score or classify multiple visualizations choices to recommend the most effective ones, as per the statistics. However, state-of-the art models rely on very large number of expensive statistics and therefore using such models on large datasets become infeasible due to prohibitively large computational time, limiting the effectiveness of such techniques to most large real world datasets. In this paper, we propose a novel reinforcement-learning (RL) based framework that takes a given Vis-Rec model and a time-budget from the user and identifies the best set of input statistics, specifically for a target dataset, that would be most effective while generating accurate enough visual insights. We show the effectiveness of our technique as it enables two state-of-the-art Vis-Rec models achieve upto 10x speedup in time-to-visualize on four large real-world datasets.

Text-to-image generation using diffusion models has seen explosive popularity owing to their ability in producing high quality images adhering to text prompts. However, production-grade diffusion model serving is a resource intensive task that not only require high-end GPUs which are expensive but also incurs considerable latency. In this paper, we introduce a technique called approximate-caching that can reduce such iterative denoising steps for an image generation based on a prompt by reusing intermediate noise states created during a prior image generation for similar prompts. Based on this idea, we present an end to end text-to-image system, Nirvana, that uses the approximate-caching with a novel cache management-policy Least Computationally Beneficial and Frequently Used (LCBFU) to provide 20% GPU compute savings, 19.8% end-to-end latency reduction and 19% dollar savings, on average, on two real production workloads. We further present an extensive characterization of real production text-to-image prompts from the perspective of caching, popularity and reuse of intermediate states in a large production environment.

Cloud services are omnipresent and critical cloud service failure is a fact of life. In order to retain customers and prevent revenue loss, it is important to provide high reliability guarantees for these services. One way to do this is by predicting outages in advance, which can help in reducing the severity as well as time to recovery. It is difficult to forecast critical failures due to the rarity of these events. Moreover, critical failures are ill-defined in terms of observable data. Our proposed method, Outage-Watch, defines critical service outages as deteriorations in the Quality of Service (QoS) captured by a set of metrics. Outage-Watch detects such outages in advance by using current system state to predict whether the QoS metrics will cross a threshold and initiate an extreme event. A mixture of Gaussian is used to model the distribution of the QoS metrics for flexibility and an extreme event regularizer helps in improving learning in tail of the distribution. An outage is predicted if the probability of any one of the QoS metrics crossing threshold changes significantly. Our evaluation on a real-world SaaS company dataset shows that Outage-Watch significantly outperforms traditional methods with an average AUC of 0.98. Additionally, Outage-Watch detects all the outages exhibiting a change in service metrics and reduces the Mean Time To Detection (MTTD) of outages by up to 88% when deployed in an enterprise cloud-service system, demonstrating efficacy of our proposed method.

Modern cloud services are prone to failures due to their complex architecture, making diagnosis a critical process. Site Reliability Engineers (SREs) spend hours leveraging multiple sources of data, including the alerts, error logs, and domain expertise through past experiences to locate the root cause(s). These experiences are documented as natural language text in outage reports for previous outages. However, utilizing the raw yet rich semi-structured information in the reports systematically is time-consuming. Structured information, on the other hand, such as alerts that are often used during fault diagnosis, is voluminous and requires expert knowledge to discern. Several strategies have been proposed to use each source of data separately for root cause analysis. In this work, we build a diagnostic service called ESRO that recommends root causes and remediation for failures by utilizing structured as well as semi-structured sources of data systematically. ESRO constructs a causal graph using alerts and a knowledge graph using outage reports, and merges them in a novel way to form a unified graph during training. A retrieval-based mechanism is then used to search the unified graph and rank the likely root causes and remediation techniques based on the alerts fired during an outage at inference time. Not only the individual alerts, but their respective importance in predicting an outage group is taken into account during recommendation. We evaluated our model on several cloud service outages of a large SaaS enterprise over the course of ∼2 years, and obtained an average improvement of 27% in rouge scores after comparing the likely root causes against the ground truth over state-of-the-art baselines. We further establish the effectiveness of ESRO through qualitative analysis on multiple real outage examples.

AI-based monitoring has become crucial for cloud-based services due to its scale. A common approach to AI-based monitoring is to detect causal relationships among service components and build a causal graph. Availability of domain information makes cloud systems even better suited for such causal detection approaches. In modern cloud systems, however, auto-scalers dynamically change the number of microservice instances, and a load-balancer manages the load on each instance. This poses a challenge for off-the-shelf causal structure detection techniques as they neither incorporate the system architectural domain information nor provide a way to model distributed compute across varying numbers of service instances. To address this, we develop CausIL, which detects a causal structure among service metrics by considering compute distributed across dynamic instances and incorporating domain knowledge derived from system architecture. Towards the application in cloud systems, CausIL estimates a causal graph using instance-specific variations in performance metrics, modeling multiple instances of a service as independent, conditional on system assumptions. Simulation study shows the efficacy of CausIL over baselines by improving graph estimation accuracy by ∼ 25% as measured by Structural Hamming Distance whereas the real-world dataset demonstrates CausIL’s applicability in deployment settings.

The ability to extract insights from large amounts of data in a timely manner is a crucial problem. Exploratory Data Analysis (EDA) is commonly used by analysts to uncover insights using a sequence of SQL commands and associated visualizations. However, in many cases, this process is carried out by non-programmers who must work within tight time constraints, such as in a marketing campaign where a marketer must quickly analyse large amounts of data to reach a target revenue. This paper presents ApproxEDA - a system that combines a natural language processing (NLP) interface for insight discovery with an underlying sample-based EDA engine. The NLP interface can convert high-level questions into contextual SQL queries of the dataset, while the backend EDA engine significantly speeds up insight discovery by selecting the most optimum sample from among many pre-created samples using various sampling strategies. We demonstrate that ApproxEDA addresses two key aspects: converting high-level NLP inputs to contextual SQL and intelligently selecting samples using a reinforcement learning agent. This protects users from diverging from their original intent of analysis, which can occur due to approximation errors in results and visualizations, while still providing optimal latency reduction through the use of samples.