Accelerating approximate aggregation queries with expensive predicates

Abstract

Researchers and industry analysts are increasingly interested in computing aggregation queries over large, unstructured datasets with selective predicates that are computed using expensive deep neural networks (DNNs). As these DNNs are expensive and because many applications can tolerate approximate answers, analysts are interested in accelerating these queries via approximations. Unfortunately, standard approximate query processing techniques to accelerate such queries are not applicable because they assume the result of the predicates are available ahead of time. Furthermore, recent work using cheap approximations (i.e., proxies) do not support aggregation queries with predicates. To accelerate aggregation queries with expensive predicates, we develop and analyze a query processing algorithm that leverages proxies (ABAE). ABAE must account for the key challenge that it may sample records that do not satisfy the predicate. To address this challenge, we first use the proxy to group records into strata so that records satisfying the predicate are ideally grouped into few strata. Given these strata, ABAE uses pilot sampling and plugin estimates to sample according to the optimal allocation. We show that ABAE converges at an optimal rate in a novel analysis of stratified sampling with draws that may not satisfy the predicate. We further show that ABAE outperforms on baselines on six real-world datasets, reducing labeling costs by up to 2.3X.

Accelerating Aggregation Queries on Unstructured Streams of Data

“…One method for showing that the bootstrap is valid is to demonstrate its asymptotic validity. The asymptotic validity of sampling with stochastic draws follows from the analysis in [28]. We can also use a standard subgaussian tail bound, but they give similar results.…”

“…We model our syntax after the Apache Flink SQL language with some minor extensions [31]. Similar to unstructured AQP systems [24,25,27], a user provides InQuest with a sampling budget, a proxy model, ⊲ duration for non-continuous queries USING proxy Figure 2: Syntax for InQuest which is based on Apache Flink SQL syntax. Users provide a statistic to compute, a dataset, a segment length defined by a tumbling window, an oracle limit per segment, and a proxy model for computing proxy scores in real-time.…”

“…Our problem setting involves a number of key challenges. Similar to some prior work [27], we do not know the correlation between the proxy model P (𝑥) and the ground-truth statistic function 𝑓 (𝑥) ahead of time. We also do not have prior knowledge of the quantities 𝜎 𝑡𝑘 and 𝑝 𝑡𝑘 .…”

“…One limitation of using DNNs for query processing is that executing large DNNs exhaustively over most real-world datasets can be prohibitively expensive. As a result, recent work has focused on accelerating DNN-based queries [4,21,[24][25][26][27]35]. One common approach involves filtering (or sub-sampling) records with inexpensive approximation models ("proxies") and then applying high-precision models ("oracles") to extract statistics of interest.…”

“…We evaluate InQuest on six real-world video and text datasets and compare it against two streaming baselines-uniform sampling and stratified sampling with fixed strata and fixed sample allocations. We also compare it against ABae [27], a state-of-the-art algorithm for the batch setting which provides answers to aggregation queries with valid confidence intervals. We show that InQuest can achieve the same root mean squared error (RMSE) as the streaming baselines with up to 5.0x fewer samples, and it can achieve up to 1.9x lower RMSE than ABae at a fixed number of oracle invocations.…”