15.1. FileSystem Datastore Architecture¶
The GeoMesa FileSystem Datastore (GeoMesa FSDS) takes advantage of the performance characteristics of modern cloud-native and distributed filesystems to scale bulk analytic queries. The FSDS is a good choice for doing bulk egress queries or large analytic jobs using frameworks such as Spark SQL and MapReduce. The FSDS differs from other datastores in that ingest and point query latencies are traded for high-throughput query performance. The FSDS pairs well with low-latency ingest and cache-based datastores systems such as HBase or Kafka to provide an optimal pairing of “hot” and “warm” storage options. This pairing is commonly known as a Lambda Architecture.
The GeoMesa FSDS consists of a few primary components:
- FileSystem - A separately managed storage system that implements the GeoMesa FileSystem API
- Partition Scheme - A stategy for laying out data on within the filesystem
- Storage Format - A defined format or encoding to store data in files
- Query Engine - A query engine or client to fulfill queries and run analytic jobs
GeoMesa FSDS can utilize any filesystem that implements the Hadoop FileSystem API. The most common filesystems used with GeoMesa FSDS are:
- HDFS - Hadoop Distributed File System
- S3 - Amazon Simple Storage
- GCS - Google Cloud Storage
- WASB - Windows Azure Blob Storage
- Local - Locally Mounted File System (e.g. local disk or NFS)
Choosing a filesystem depends generally on cost and performance requirements. One thing to note is that S3, GCS, and WASB are all “cloud-native” storage meaning that they are built into Amazon, Google, and Microsoft Azure cloud platforms. These cloud-native filesystems are scaled separately from the compute nodes which generally provides a more cost efficient storage solution. Compared to HDFS, their price per GigaByte of storage is lower but their latency is higher. They also have the ability to persist data after you turn off all your compute nodes.
Any of the filesystems mentioned about are good choices for the FSDS. If you have more questions about making a choice contact the GeoMesa team
15.1.2. Partition Schemes¶
The partition scheme defines how data is stored within the filesystem. The scheme is important because it defines how the data is queried. Most OLAP queries in GeoMesa contain a date range and geometric predicate and the partition scheme can aid in finding the files satisfy the query. There are two main partition schmes used with GeoMesa FSDS:
- Date - partition data by a Date attribute
- Geometry (Z2) - partition data by its geometric coordinates using a Z2 space-filling curve
- Combined Date and Geometry - partition data using a combined data-time scheme
The partition scheme must be provided at ingest time. Examples of common schemes are:
- Hourly - Store a file for each hour of each day
- Daily - Store a single file for each day of the year
- Day with Z2 - Store a file each day for each region of the world (at some precision of geometry)
GeoMesa FSDS stores metadata about partitions and data files, to avoid having to interrogate the filesystem
repeatedly. When a new data file is added or removed, an associated metadata file will be created to track
the operation. These files are stored in a folder named
metadata under the root path for the FSDS.
If the number of metadata files grows too large, they may be reduced down by using the
manage-metadata compact command-line functions, and/or manually moved into subfolders.
15.1.4. Storage Formats¶
- Apache Parquet - Apache Parquet is the leading interoperable columnar format in the Hadoop ecosystem. It provides efficient compression, storage, and query of structured data.
- Apache ORC - Apache ORC is a self-describing type-aware columnar file format designed for Hadoop workloads. It is optimized for large streaming reads, but with integrated support for finding required rows quickly.
- Converter Storage - The converter storage format is a synthetic format which allows you to overlay a GeoMesa converter on top of a filesystem using a defined partition scheme. This allows you to utilize existing data storage layouts of data stored in JSON, CSV, TSV, Avro, or other formats. Converters are pluggable allowing users to expose their own custom storage formats if desired. Converter storage is a read-only format.