Optimizing Geospatial Operations with Server-side Programming in HBase and Accumulo
Download as PPTX, PDF•4 likes•1,591 views
The document outlines the optimization of geospatial operations using server-side programming in HBase and Accumulo, detailing the benefits of GeoMesa for managing and analyzing spatio-temporal data at scale. Key topics include indexing with space filling curves, filtering and aggregation techniques, and the implementation of server-side optimizations to enhance performance. It also covers practical applications, integration with frameworks like MapReduce and Spark, and the support for interactive data discovery using notebook servers.
![GeoMesa Spark Example 1: Time SeriesStep 1: Get an RDD[SimpleFeature]
Step 2: Calculate the time series
Step 3: Plot the time series in R.](https://app.smartcontext.ai/default/https://image.slidesharecdn.com/tues1030amoptimizinggeospatialoperationswithserver-sideprogramminginhbaseandaccumulo-190612045313/85/Optimizing-Geospatial-Operations-with-Server-side-Programming-in-HBase-and-Accumulo-58-320.jpg)
Optimizing Geospatial Operations with Server-side Programming in HBase and Accumulo
- 1. Optimizing Geospatial Operations with Server-side Programming in HBase and Accumulo James Hughes, CCRi
- 2. James Hughes ● CCRi’s Director of Open Source Programs ● Working in geospatial software on the JVM for the last 7 years ● GeoMesa core committer / product owner ● SFCurve project lead ● JTS committer ● Contributor to GeoTools and GeoServer
- 3. ● Background / Warm-up / What we are talking about ○ What is GeoMesa? ○ Quick Demo ● General Implementation Details ○ Indexing on Accumulo/HBase with Space Filling Curves ○ Filtering/transforming ■ Applying secondary filters ■ Changing output (projections / format changes) ○ Aggregations ■ Heatmaps ■ Stats ● Database specifics ○ Accumulo Implementation details ○ HBase Implementation details Talk outline
- 4. Motivation ● What is geospatial? ● IoT based data examples?
- 5. Spatial Data Types Points Locations Events Instantaneous Positions Lines Road networks Voyages Trips Trajectories Polygons Administrative Regions Airspaces
- 7. Topology Operations 7 Algorithms ● Convex Hull ● Buffer ● Validation ● Dissolve ● Polygonization ● Simplification ● Triangulation ● Voronoi ● Linear Referencing ● and more...
- 8. GeoMesa ● GeoMesa Overview
- 9. What is GeoMesa? A suite of tools for streaming, persisting, managing, and analyzing spatio- temporal data at scale
- 10. What is GeoMesa? A suite of tools for streaming, persisting, managing, and analyzing spatio- temporal data at scale
- 11. What is GeoMesa? A suite of tools for streaming, persisting, managing, and analyzing spatio-temporal data at scale
- 12. What is GeoMesa? A suite of tools for streaming, persisting, managing, and analyzing spatio- temporal data at scale
- 13. What is GeoMesa? A suite of tools for streaming, persisting, managing, and analyzing spatio- temporal data at scale
- 15. Live Demo! ● Filtering by spatio-temporal constraints ● Filtering by attributes ● Aggregations ● Transformations
- 16. Indexing Geospatial Data ● Key Design using Space Filling Curves
- 17. ● Goal: Index 2+ dimensional data ● Approach: Use Space Filling Curves Space Filling Curves (in one slide!)
- 18. ● Goal: Index 2+ dimensional data ● Approach: Use Space Filling Curves ● First, ‘grid’ the data space into bins. Space Filling Curves (in one slide!)
- 19. ● Goal: Index 2+ dimensional data ● Approach: Use Space Filling Curves ● First, ‘grid’ the data space into bins. ● Next, order the grid cells with a space filling curve. ○ Label the grid cells by the order that the curve visits the them. ○ Associate the data in that grid cell with a byte representation of the label. Space Filling Curves (in one slide!)
- 20. ● Goal: Index 2+ dimensional data ● Approach: Use Space Filling Curves ● First, ‘grid’ the data space into bins. ● Next, order the grid cells with a space filling curve. ○ Label the grid cells by the order that the curve visits the them. ○ Associate the data in that grid cell with a byte representation of the label. ● We prefer “good” space filling curves: ○ Want recursive curves and locality. Space Filling Curves (in one slide!)
- 21. ● Goal: Index 2+ dimensional data ● Approach: Use Space Filling Curves ● First, ‘grid’ the data space into bins. ● Next, order the grid cells with a space filling curve. ○ Label the grid cells by the order that the curve visits the them. ○ Associate the data in that grid cell with a byte representation of the label. ● We prefer “good” space filling curves: ○ Want recursive curves and locality. ● Space filling curves have higher dimensional analogs. Space Filling Curves (in one slide!)
- 22. To query for points in the grey rectangle, the query planner enumerates a collection of index ranges which cover the area. Note: Most queries won’t line up perfectly with the gridding strategy. Further filtering can be run on the tablet/region servers (next section) or we can return ‘loose’ bounding box results (likely more quickly). Query planning with Space Filling Curves
- 23. Server-Side Optimizations Filtering and transforming records ● Pushing down data filters ○ Z2/Z3 filter ○ CQL Filters ● Projections
- 24. Filtering and transforming records overview Using Accumulo iterators and HBase filters, it is possible to filter and map over the key-values pairs scanned. This will let us apply fine-grained spatial filtering, filter by secondary predicates, and implement projections.
- 25. Pushing down filters Let’s consider a query for tankers which are inside a bounding box for a given time period. GeoMesa’s Z3 index is designed to provide a set of key ranges to scan which will cover the spatio-temporal range. Additional information such as the vessel type is part of the value. Using server-side programming, we can teach Accumulo and HBase how to understand the records and filter out undesirable records. This reduces network traffic and distributes the work.
- 26. Projection To handle projections in a query, Accumulo Iterators and HBase Filters can change the returned key-value pairs. Changing the key is a bad idea. Changing the value allows for GeoMesa to return a subset of the columns that a user is requesting.
- 27. GeoMesa Server-Side Filters ● Z2/Z3 filter ○ Scan ranges are not decomposed enough to be very accurate - fast bit-wise comparisons on the row key to filter out-of-bounds data ● CQL/Transform filter ○ If a predicate is not handled by the scan ranges or Z filters, then slower GeoTools CQL filters are applied to the serialized SimpleFeature in the row value ○ Relational projections (transforms) applied to reduce the amount of data sent back ● Other specialized filters ○ Age-off for expiring rows based on a SimpleFeature attribute ○ Attribute-key-value for populating a partial SimpleFeature with an attribute value from the row ○ Visibility filter for merging columns into a SimpleFeature when using attribute-level visibilities
- 28. Server-Side Optimizations Aggregations ● Generating heatmaps ● Descriptive Stats ● Arrow format
- 29. Aggregations Using Accumulo Iterators and HBase coprocessors, it is possible to aggregate multiple key-value pairs into one response. Effectively, this lets one implement map and reduce algorithms. These aggregations include computing heatmaps, stats, and custom data formats. The ability to aggregate data can be composed with filtering and projections.
- 30. GeoMesa Aggregation Abstractions Aggregation logic is implemented in a shared module, based on a lifecycle of 1. Initialization 2. observing some number of features 3. aggregating a result. This paradigm is easily adapted to the specific implementations required by Accumulo and HBase. Notably, all the algorithms we describe work in a single pass over the data.
- 31. GeoMesa Aggregation Abstractions The main logic is contained in the AggregatingScan class:
- 32. Visualization Example: Heatmaps Without powerful visualization options, big data is big nonsense. Consider this view of shipping in the Mediterranean sea
- 33. Visualization Example: Heatmaps Without powerful visualization options, big data is big nonsense. Consider this view of shipping in the Mediterranean sea
- 34. Generating Heatmaps Heatmaps are implemented in DensityScan. For the scan, we set up a 2D grid array representing the pixels to be displayed. On the region/tablet servers, each feature increments the count of any cells intersecting its geometry. The resulting grid is returned as a serialized array of 64-bit integers, minimizing the data transfer back to the client. The client process merges the grids from each scan range, then normalizes the data to produce an image. Since less data is transmitted, heatmaps are generally faster.
- 35. Statistical Queries We support a flexible stats API that includes counts, min/max values, enumerations, top-k (StreamSummary), frequency (CountMinSketch), histograms and descriptive statistics. We use well-known streaming algorithms backed by data structures that can be serialized and merged together. Statistical queries are implemented in StatsScan. On the region/tablet servers, we set up the data structure and then add each feature as we scan. The client receives the serialized stats, merges them together, and displays them as either JSON or a Stat instance that can be accessed programmatically.
- 36. Arrow Format Apache Arrow is a columnar, in-memory data format that GeoMesa supports as an output type. In particular, it can be used to drive complex in-browser visualizations. Arrow scans are implemented in ArrowScan. With Arrow, the data returned from the region/tablet servers is similar in size to a normal query. However, the processing required to generate Arrow files can be distributed across the cluster instead of being done in the client. As we scan, each feature is added to an in-memory Arrow vector. When we hit the configured batch size, the current vector is serialized into the Arrow IPC format and sent back to the client. All the client needs to do is to create a header and then concatenate the batches into a single response.
- 37. Server-Side Optimizations Data ● Row Values ● Tables/compactions
- 38. Row Values Our first approach was to store each SimpleFeature attribute in a separate column. However, this proved to be slow to scan. Even when skipping columns for projections, they are still loaded off disk. Column groups seem promising, but they kill performance if you query more than one.
- 39. Row Values Our second (and current) approach is to store the entire serialized SimpleFeature in one column. Further optimizations: ● Lazy deserialization - SimpleFeature implementation that wraps the row value and only deserializes each attribute as needed ● Feature ID is already stored in the row key to prevent row collisions, so don’t also store it in the row value ● Use BSON for JSON serialization, along with JsonPath extractors ● Support for TWKB geometry serialization to save space
- 40. Tables/Compactions When dealing with streaming data sources, continuously writing data to a table will cause a lot of compactions. Table partitioning can mitigate this by creating a new table per time period (e.g. day/week), extracted from the SimpleFeature. Generally only the most recent table(s) will be compacted. For frequent updates to existing features, the GeoMesa Lambda store uses Kafka as a medium-term cache before persisting to the key-value store. This reduces the cluster load significantly.
- 41. Accumulo Server Side Programming ● Accumulo Iterator Review ● GeoMesa’s Accumulo iteraors
- 42. “Iterators provide a modular mechanism for adding functionality to be executed by TabletServers when scanning or compacting data. This allows users to efficiently summarize, filter, and aggregate data.” -- Accumulo 1.7 documentation Part of the modularity is that the iterators can be stacked: t the output of one can be wired into the next. Example: The first iterator might read from disk, the second could filter with Authorizations, and a final iterator could filter by column family. Other notes: ● Iterators provided a sorted view of the key/values. Accumulo Iterators
- 43. A request to GeoMesa consists of two broad pieces: 1. A filter restricting the data to act on, e.g.: a. Records in Maryland with ‘Accumulo’ in the text field. b. Records during the first week of 2016. 2. A request for ‘how’ to return the data, e.g.: a. Return the full records b. Return a subset of the record (either a projection or ‘bin’ file format) c. Return a histogram d. Return a heatmap / kernel density Generally, a filter can be handled partially by selecting which ranges to scan; the remainder can be handled by an Iterator. Modifications to selected data can also be handled by a GeoMesa Iterator. GeoMesa Data Requests
- 44. The first pass of GeoMesa iterators separated concerns into separate iterators. The GeoMesa query planner assembled a stack of iterators to achieve the desired result. Initial GeoMesa Iterator design Image from “Spatio-temporal Indexing in Non-relational Distributed Databases” by Anthony Fox, Chris Eichelberger, James Hughes, Skylar Lyon
- 45. The key benefit to having decomposed iterators is that they are easier to understand and re-mix. In terms of performance, each one needs to understand the bytes in the Key and Value. In many cases, this will lead to additional serialization/deserialization. Now, we prefer to write Iterators which handle transforming the underlying data into what the client code is expecting in one go. Second GeoMesa Iterator design
- 46. 1. Using fewer iterators in the stack can be beneficial 2. Using lazy evaluation / deserialization for filtering Values can power speed improvements. 3. Iterators take in Sorted Keys + Values and *must* produce Sorted Keys and Values. Lessons learned about Iterators
- 47. HBase Server Side Programming ● HBase Filter and Coprocessor Review ● GeoMesa HBase Filter ● GeoMesa HBase Coprocessor
- 48. HBase Filter Info HBase filters are restricted to the ability to skip/include rows, and to transform a row before returning it. Anything more complicated requires a Coprocessor. In contrast to Accumulo, where iterators are configured with a map of options, HBase requires custom serialization code for each filter implementation.
- 49. HBase Filter Info The main GeoMesa filters are: ● org.locationtech.geomesa.hbase.filters.CqlTransformFilter ○ Filters rows based on GeoTools CQL ○ Transforms rows based on relational projections ● org.locationtech.geomesa.hbase.filters.Z2HBaseFilter ○ Compares Z-values against the row key ● org.locationtech.geomesa.hbase.filters.Z3HBaseFilter ○ Compares Z-values against the row key
- 50. HBase Coprocessor Info Coprocessors are not trivial to implement or invoke, and can starve your cluster if it is not configured correctly. GeoMesa implements a harness to invoke a coprocessor, receive the results, and handle any errors: ● org.locationtech.geomesa.hbase.coprocessor.GeoMesaCoprocessor An adapter layer links the common aggregating code to the coprocessor API: ● org.locationtech.geomesa.hbase.coprocessor.aggregators.HBaseAggregato r
- 51. HBase Coprocessor Info GeoMesa defines a single Protobuf coprocessor endpoint, modeled around the Accumulo iterator lifecycle. The aggregator class and a map of options are passed to the endpoint. Each aggregating scan requires a trivial adapter implementation: ● HBaseDensityAggregator ● HBaseStatsAggregator ● HBaseArrowAggregator
- 52. Thanks! James Hughes ● jhughes@ccri.com ● http://geomesa.org ● http://gitter.im/locationtech/geomesa
- 53. Backup Slides Integration with MapReduce / Spark ● GeoMesa + Spark Setup ● GeoMesa + Spark Analytics ● GeoMesa powered notebooks (Jupyter and Zeppelin)
- 54. Using Accumulo Iterators, we’ve seen how one can easily perform simple ‘MapReduce’ style jobs without needing more infrastructure. NB: Those tasks are limited. One can filter inputs, transform/map records and aggregate partial results on each tablet server. To implement more complex processes, we look to MapReduce and Spark. GeoMesa MapReduce and Spark Support
- 55. Using Accumulo Iterators, we’ve seen how one can easily perform simple ‘MapReduce’ style jobs without needing more infrastructure. NB: Those tasks are limited. One can filter inputs, transform/map records and aggregate partial results on each tablet server. To implement more complex processes, we look to MapReduce and Spark. Accumulo Implements the MapReduce InputFormat interface. GeoMesa MapReduce and Spark Support
- 56. Using Accumulo Iterators, we’ve seen how one can easily perform simple ‘MapReduce’ style jobs without needing more infrastructure. NB: Those tasks are limited. One can filter inputs, transform/map records and aggregate partial results on each tablet server. To implement more complex processes, we look to MapReduce and Spark. Accumulo Implements the MapReduce InputFormat interface. Spark provides a way to change InputFormats into RDDs. GeoMesa MapReduce and Spark Support
- 57. Using Accumulo Iterators, we’ve seen how one can easily perform simple ‘MapReduce’ style jobs without needing more infrastructure. NB: Those tasks are limited. One can filter inputs, transform/map records and aggregate partial results on each tablet server. To implement more complex processes, we look to MapReduce and Spark. Accumulo Implements the MapReduce InputFormat interface. GeoMesa MapReduce and Spark Support
- 58. GeoMesa Spark Example 1: Time SeriesStep 1: Get an RDD[SimpleFeature] Step 2: Calculate the time series Step 3: Plot the time series in R.
- 59. Using one dataset (country boundaries) to group another (here, GDELT) is effectively a join. Our summer intern, Atallah, worked out the details of doing this analysis in Spark and created a tutorial and blog post. This picture shows ‘stability’ of a region from GDELT Goldstein values GeoMesa Spark Example 2: Aggregating by Regions http://www.ccri.com/2016/08/17/new-geomesa-tutorial-aggregating-visualizing-data/ http://www.geomesa.org/documentation/tutorials/shallow-join.html
- 60. GeoMesa Spark Example 3: Aggregating Tweets about #traffic Virginia Polygon CQL GeoMesa RDD Aggregate by County Calculate ratio of #traffic Store back to GeoMesa
- 61. GeoMesa Spark Example 3: Aggregating Tweets about #traffic #traffic by Virginia county Darker blue has a higher count
- 62. Interactive Data Discovery at Scale in GeoMesa Notebooks Writing (and debugging!) MapReduce / Spark jobs is slow and requires expertise. A long development cycle for an analytic saps energy and creativity. The answer to both is interactive ‘notebook’ servers like Apache Zeppelin and Jupyter (formerly iPython Notebook).
- 63. Interactive Data Discovery at Scale in GeoMesa Notebooks Writing (and debugging!) MapReduce / Spark jobs is slow and requires expertise. A long development cycle for an analytic saps energy and creativity. The answer to both is interactive ‘notebook’ servers like Apache Zeppelin and Jupyter There are two big things to work out: 1. Getting the right libraries on the classpath. 2. Wiring up visualizations.
- 64. Interactive Data Discovery at Scale in GeoMesa Notebooks GeoMesa Notebook Roadmap: ● Improved JavaScript integration ● D3.js and other visualization libraries ● OpenLayers and Leaflet ● Python Bindings
- 65. Backup Slides Indexing non-point geometries
- 66. Most approaches to indexing non-point geometries involve covering the geometry with a number of grid cells and storing a copy with each index. This means that the client has to deduplicate results which is expensive. Indexing non-point geometries: XZ Index
- 67. Most approaches to indexing non-point geometries involve covering the geometry with a number of grid cells and storing a copy with each index. This means that the client has to deduplicate results which is expensive. Böhm, Klump, and Kriegel describe an indexing strategy allows such geometries to be stored once. GeoMesa has implemented this strategy in XZ2 (spatial-only) and XZ3 (spatio-temporal) tables. The key is to store data by resolution, separate geometries by size, and then index them by their lower left corner. This does require consideration on the query planning side, but avoiding deduplication is worth the trade-off. Indexing non-point geometries: XZ Index For more details, see Böhm, Klump, and Kriegel. “XZ-ordering: a space-filling curve for objects with spatial extension.” 6th. Int. Symposium on Large Spatial Databases (SSD), 1999, Hong Kong, China. (http://www.dbs.ifi.lmu.de/Publikationen/Boehm/Ordering_99.pdf)
- 69. Here the viewport is used as the spatial bounds for the query. The time range is a 12 hour period on Monday. Query by bounding box
- 70. Query by polygon Here we further restrict the query region by an arbitrary polygon
- 71. Query by polygon and vessel type Here, we have added a clause to restrict to cargo vessels
- 72. Query by polygon and vessel type (heatmap) Heatmaps can be generated
- 73. Query by polygon and vessel type (Apache Arrow format) Data can be returned in a number of formats. The Apache Arrow format allows for rapid access in JavaScript. Here, points are colored by callsign.
- 74. Query by polygon and vessel type (Apache Arrow format) Apache Arrow allows for in browser data exploration. This histogram shows callsigns grouped by country. Selections in the histogram can influence the map.