On-Demand Recordings from PrestoCon’s, Webinars, Meetups, and more
In this session, we will dive deeper into the Native Hudi connector, walkthrough the various checkpoints where the query execution moves from Presto to Hudi library, and then discuss new features that are currently being developed.
This month we’ll have Yanbing Zhang, software engineer at Bytedance, share more about their work on the Apache Hudi Connector for Presto and how they’re using it at Bytedance.
Vladimir Rodionov, founder of Carrot Cache will present the Velociraptor – the next evolution of PrestoDB hierarchical caching framework RaptorX. Velociraptor enables efficient data and meta-data caching well beyond RaptorX limits in terms of number of data files (multi-billions), number of table partitions (multi-millions) and number of table columns (multi-thousands). Velociraptor replaces all five RaptorX caches (Hive meta-data, file list, query result fragments, ORC/Parquet meta-data and data I/O) with a scalable solution, based on Carrot Cache, which does not pollute JVM heap memory, does not affect Java Garbage Collector, keeps all data and meta-data off Java heap memory or on disk and can scale well beyond server’s physical RAM limit. Velociraptor supports server restart, by quickly saving and loading data to/from disk for automatic cache warm up.
Danny Chan & Sagar Sumit, Onehouse – In this talk, attendees will walk away with: – The current challenges of analytics on transactional data systems with data streams at scale – How the Hudi unlocks incremental processing on the lake – How Presto allows ad-hoc queries that support data exploration on Flink data – How you can leverage Flink, Hudi and Presto to build incremental materialized views
Sivabalan Narayanan of Onehouse shares more about how Apache Hudi brought transactions, incremental processing on top of data lakes, which are deemed as the foundational pillars for Lakehouse architecture. In this session, we will discuss Apache Hudi and how it fills the key technology gaps in the modern data architecture. Viewed from a data engineering lens, Hudi also plays a key unifying role between the batch and stream processing worlds realized by incremental processing model. We will take a look at the capabilities of native Hudi connector in Presto. We will dive deep into this connector, covering the key optimizations and features it unblocks. Presto users could now leverage the metadata table for optimized file listing and avoid large number of list operations in cloud storages. We will look at how we can improve the query latency in Presto using advanced data skipping methodogies employed with multi-modal sub-system with Hudi.
In this talk we are going to introduce Hudi, discuss different table/query types and how Hudi integrates with Presto to support these queries. We like to share our experience on how this integration has evolved over time and also discuss upcoming file listing and query planning improvements in Presto Hudi queries.
Apache Hudi is a rich platform to build self-managing, exabyte-scale data lakes, optimized for incremental as well as regular batch processing. Hudi tables can be seamlessly synced to Hive metastore, which unlocks the powerful capabilities of Presto engine via the Hive connector. Presto-Hudi integration is over five years old. What started as simply fetching splits using a custom input format for a Hudi Copy-On-Write table has evolved into snapshot querying of Merge-On-Read tables and using Hudi’s internal metadata table to boost query performance. In this session, we trace that journey and discuss in detail the recent developments that have made this integration stronger not only in terms of usability but also performance. We discuss the additional features that come with the brand new presto-hudi connector, such as multi-modal index and data skipping for better query performance.
RaptorX, an umbrella project presented in PrestoCon Day in March, enabled the Presto interactive fleet in Facebook to reduce latency by 10x, based on a set of architectural improvements and optimizations with hierarchical caching. This presentation provides an update on the follow-up enhancement. Bin Fan from Alluxio will talk about the exploration of a probabilistic algorithm in Alluxio caching to estimate cache working set and the implementation of shadow cache Ke Wang from Facebook will talk about how shadow cache is used to understand the system bottleneck for better resource allocation and query routing decisions. She will also cover a recent improvement in collecting and aggregating per-query runtime statistics on the Presto engine to better understand the time breakdown, resource usage breakdown and cache hit rate on a per-query basis, which can help identify areas of improvement.
RaptorX is an internal project name aiming to boost query latency significantly beyond what vanilla Presto is capable of. For this session, we introduce the hierarchical cache work including Alluxio data cache, fragment result cache, etc. Cache is the key building block for RaptorX. With the support of the cache, we are able to boost query performance by 10X. This new architecture can beat performance oriented connectors like Raptor with the added benefit of continuing to work with disaggregated storage.
The Presto Ecosystem – Moderated by Dipti Borkar, Ahana; Maxime Beauchemin, Preset; Vinoth Chandar, Apache Hudi; Kishore Gopalakrishna, Apache Pinot & James Sun, Facebook, Inc.
Apache Hudi is a data lake platform that supercharges data lakes. Originally created at Uber, Hudi provides various ways to strike trade-offs between ingestion speed and query performance by supporting user defined partitioners, automatic file sizing which are favorable to query performance. Hudi integrates with PrestoDB to make this data available for queries. During ingestion, data is typically co-located based on arrival time. However, query engines perform better when the data frequently queried is co-located together, which may be different from arrival time order. We will discuss a new framework called “data clustering” to make data lakes adaptable to query patterns, thereby improving query latencies. Finally, we will discuss future work to support improving data locality using custom bucketing of data during ingestion, avoiding some of the rewrite costs.