Applications within Google Google is not affiliated or endorsed by any of these companies. Apache HBase, Apache Cassandra and Apache Accumulo are trademarks are of The Apache Software Foundation. Hypertable is the trademark of Hypertable Inc.
decided to build a database service that could scale linearly across thousands and thousands of commodity servers ◦ Systems will fail, retain performance at scale • Leave the traditional relational model to achieve goals • The first generation was about: ◦ Prototyping and build the service to do its first scaling ◦ Migrate initial applications to Bigtable ◦ Invent replication and first multi-tenant version of Bigtable ◦ Painful rediscovery
- now web serving as well ◦ Making it very low latency and bringing in the 99th % of requests [this is a hard problem] • Perfecting the Bigtable service ◦ What is that: a multi-tenant shared service model for a single database on a common set of resources [this is a hard problem] ▪ Spikes in CPU happen quickly and reacting to abusive usage is difficult to do effectively ▪ Hard-capping leaves resources on the table, and you lose the agility and efficiency you were looking for
we think we can get a pretty high cache hit rate with a modest increase in memory • Mixed media clusters - mixture of SSD + HDD storage and an ability to specify an affinity • Tabletserver failure - Target is recovery in 1 second or less rather than 10s of seconds or minutes = appears to customer as latency if at all • Effortless Bigtable replication either in multiple zones for higher availability or across the world for better latency
a fully-managed service, simplifying operations and management of applications • Cloud Bigtable allows developers to quickly build applications to an industry standard API with no need to focus on infrastructure • Simple pricing model with serve resources and storage resources separated • High performance, and low latency, and low cost, and little to no configuration
written to Cloud Bigtable through a RESTful or RPC-based data service layer. Typically this will be to serve data to applications, dashboards and other microservices. Streaming Data can be streamed in (written event by event) through a variety of popular stream processing frameworks. Batch Processing Data can be read from and written to Cloud Bigtable through batch processing systems (either MapReduce based or analytical). Often, summarized or newly calculated data is written back to Cloud Bigtable or to a downstream database. Review Typical Access Patterns
1.0+ API/Client • While HBase is a separate system from Bigtable we have close ties to the community • We like the community - lots of voices, moving together, reps from many major tech giants, very widely adopted • Semantics and operations are very similar ◦ Want it to be easy to understand, transition to, develop against • Release tools that work with Cloud Bigtable and HBase and vice versa ◦ Grow the whole community so that all benefit
and change the serving resources with a single button with a single per-hour pricing ◦ What are Bigtable nodes? ◦ This is just the raw compute power that makes up the serve path - separate from the persisted storage tier • You’re billed separately for the amount of storage you use of whatever medium you choose (SSD or HDD) • This makes it super simple to plan for your workload and understand what your costs are
deliver up to 10,000 QPS and 10 MB/s of throughput Cost per hour Minimum number of nodes per cluster $0.65 3 Storage SSD storage (GB/mo) HDD storage (GB/mo) (coming soon) $0.17 $0.026 On creation of a Bigtable cluster, customers provision throughput for their workload in the form of Bigtable nodes. Storage is charged on a per-use basis.
Media User engagement, clickstream analysis, real-time adaptive content Internet of Things Sensor data dashboards and anomaly detection Telecommunications Sampled traffic patterns, metric collection and reporting Energy Oil well sensors, anomaly detection, predictive modeling Biomedical Genomics sequencing data analysis Cloud Bigtable Use Cases