Article Source

• Title: Apache Tajo Internal
• Authors: Jihoon Son

Apache Tajo Internal

Hadoop has been regarded as a de-facto standard storage for web-scale data. Since SQL is popularly used for analytic works, the needs for SQL processing on Hadoop has been increased. As a result, many SQL-on-Hadoop systems have been proposed for the last few years.

Apache Tajo is one of the most promising SQL-on-Hadoop systems. It is designed to converge the technologies of Hadoop and parallel databases. As a result, it can provide scalable and efficient SQL processing. In this article, I’ll describe the internal architecture and interesting features of Tajo.

Architectural characteristics

Tajo has its own query engine based on the traditional shared-nothing architecture. This query engine adopts promising features from both of MapReduce and parallel databases.

Sequential stage execution & native execution engine

Tajo executes consecutive stages sequentially during execution of a query. This is similar to that the reduce stage is executed after the map stage is completed in MapReduce. For example, let’s consider a query that joins two relations. This query is executed in two stages in Tajo. Both relations are scanned in the first stage, and then read tuples are joined in the second stage.

In each stage, workers execute the query using a database-inspired native engine. This engine exploits pipelined operator execution of databases. In addition, advanced database techniques are adopted for operator execution.

The query processing in Tajo has interesting features as follows.

• By exploiting sequential stage execution, the entire resources of the cluster can be utilized for each stage. Thus, Tajo can work effectively even when the remaining cluster resources are very small.
• Different to Hive or Shark, parallel execution is performed by the native engine rather than other distributed processing frameworks. In this architecture, query optimizer can be more tightly integrated with query executor. Thus, more fine-grained optimization can be achieved.

Dynamic load balancing

A stage is divided into many small tasks. During execution of a stage, a query master progressively assigns tasks to workers according to their status, such as remaining resources, average throughput, etc. As a result, load balancing can be achieved according to dynamically changing cluster environments during query execution.

Fault tolerance

Tajo provides MapReduce-like fault-tolerance strategy. A Tajo master continuously gathers worker status information. If some workers do not send their status, they are marked as failed workers and excluded from query execution. Even when a worker fails in the middle of a query, tasks assigned to that worker are simply re-assigned to healthy workers.

Additional features

• Dynamic query optimization. Tajo dynamically refines the execution plan of running queries. Please note that consecutive stage are sequentially executed. During execution of a stage, some statistics such as partition size, record counts, and histograms, can be collected. After that, at the beginning of the next stage, the operator type or the degree of parallelism of the stage can be chosen based on the collected statistics.
• Columnar storage support. Parquet is inspired from Google’s Dremel, and one of the most popular columnar file formats. Tajo can provide more efficient query processing by using Parquet.

Future

Tajo has been rapidly developed in the last few years, and there are more interesting features that will be developed.

• Improved query scheduler. Currently, Tajo only supports the FIFO-style query scheduling. This is not appropriate for most real applications. In this issue, a Sparrow-like improved query scheduler have been being implemented.
• Vectorized execution & run-time code generation. Columnar store and related techniques have been interested for many years, because they greatly improve the processing performance of analytic queries. There are several ongoing issues to adopt these techniques to the Tajo query engine. This slide will give more information.
• Index support. Indexing is a representative technique to improve the performance by reducing I/O overheads. Tajo is also considering the index support. Here is a related issue.

In addition to the above features, there are a lot of ongoing interesting issues for more stable, more scalable, and more efficient query execution.