Building Scalable IoT Apps (QCon S-F)
Download as PPTX, PDF•5 likes•1,285 views
This document discusses building scalable IoT applications using open source technologies. It begins by providing an overview of the growth of the IoT market and connected devices. It then discusses challenges with traditional "data lake" architectures for IoT data due to the high volume, velocity, and variety of IoT data. The document proposes an architecture combining stream processing for real-time data with analytics on both real-time and stored data. It discusses data access patterns and storage requirements for different types of IoT data. Finally, it provides an overview of open source technologies that can be used to build scalable IoT applications.
Building Scalable IoT Apps (QCon S-F)
- 1. Building scalable IoT apps using OSS technologies Pavel Hardak Basho Technologies Disclaimer: some of the opinions expressed here are mine and might not fully agree with those of
- 2. IOT & INDUSTRY VERTICALS
- 3. IoT market - growth prediction Number of connected “things” • 2016 – about 6.4 B • 30% YoY growth, 5.5M activations per day • 2020 – about 21 B “By 2020 more than half of new major business processes and systems will incorporate some element of Internet of Things”
- 4. Reality Check - let us get a second opinion
- 6. IoT Project Plan • Investigate those “things” and figure out • What protocols they support (CoAP, MQTT, HTTP, …) • What data they generate (temperature, humidity, location, speed, ...) • Collect this data in our data center • Implement protocols and parsing routines • Store into persistent storage (“Data Lake” architecture) • Once stored in Data Lake • Analyze, summarize, “slice and dice” • Predict, discover insights • Declare a victory – make profit & go for IPO
- 7. Data Lake IoT Devices SQL Apps & AnalyticsMQTT, CoAP and HTTP REFERENCE ARCHITECTURE (?) Not so fast, my friend.
- 8. What is wrong with “Data Lake” for IoT ?
- 13. Auto Insurance - Micro Case Study • One of top 5 auto insurance companies in USA, appears in Fortune-500 list • More than $10B in annual revenue, above $15B in assets • About 20,000 employees and 50,000 insurance agents • More than 19 million individual policies across all 50 states
- 14. How this “rating info” influences your payment ? • Garaging Zip – what neighborhood is the car parked when it is not used? There is a high correlation between Zip code and the probability of car being stolen or vandalized. • Current and Previous Annual Mileage – if the insured drives for longer distances, it leads to the higher probability of road accidents or car malfunctions. • Vehicle Usage – do you use your car for work or pleasure? Are you commuter, student, stay-at-home parent or Uber driver? Depending on your usage, the company will calculate the risk and adjust the rate. • Years of Driving Experience – young drivers are put into higher risk categories, where older people are considered safer drivers due to more time behind the wheel. Note - average young driver vs. average experienced driver.
- 16. Sampling Frequency and Dataset Size • Mileage • From one sample per year to 52 (weekly) or 365 (daily) • Better - let us do hourly to “see” the car usage (commuter, …) • Location (used to be “Garaging Zip”) • From one sample per year to 365 (daily) • Better - hourly, allows to learn when car is parked for several hours • New factors for rating algorithm based on weekly summaries • Hard brakes, hard accelerations, going above the speed limit, … • Amount of time series data to be stored and analyzed • Grows by factor of 365x, then by another 24x = 8760x Each week – at least 50x more data than the whole previous year.
- 18. What is different special about IoT? It is about the “things”… and more.
- 21. IoT Data Categories Category Description Metadata & Profiles Devices Device info (model, SN, firmware, sensors, ..), configuration, owner, … Users Personal info, preferences, billing info, registered devices, … Time Series Ingested (“Raw”) Measurements, statuses and events from devices. Aggregated (“Derived”) Calculated data - from devices & profiles • Rollups – aggregate metrics from low resolution to higher ones (min - hour – day) using min, max, avg, ... • Aggregations – aggregate measurements, configuration and profiles (model, region, …) over time ranges
- 22. IOT - NETWORKING TECHNOLOGIES
- 23. NETWORK WISH LIST • Extreme Reliability • Guaranteed Delivery • End-to-End Low Latency • Quality of Service • Engineered Topology • Committed Bandwidth (CIR) • Fiber-optic network • Dedicated Channel • Strong Signal • Interference and Crosstalk Resistant • High SNR (Signal to Noise Ratio) • Very Low BER (Bit Error Rate)
- 24. REALITY CHECK - LET US LOOK AGAIN
- 25. IOT & NETWORK - REALITY • Wireless technologies • Shared transmission media • Limited bandwidth • Mesh or Ad-hoc Topology • Possible signals interference • Mis-ordered or lost packets • Low cost hardware components • Low power radio transmitters • Very small antennas • “Custom-made” firmware • Constrained Application Protocol (CoAP) • “Best Effort” QoS (“shoot and forget”)
- 26. IoT is “Big Data” - by definition. Actually, lots and lots of Big Data.
- 27. Five “V”s IoT data Velocity Torrent of small writes (sensors). Reads – millions of low-latency queries: user and device profiles, range queries for TS data (slices). Stream of updates (profiles) - beware of conflicts.
- 28. Five “V”s IoT data Velocity Torrent of small writes (sensors). Reads – millions of low-latency queries, user and device profiles, range queries for TS data (slices). Stream of updates (profiles) - beware of conflicts. Variety Sensors data (time series), users and devices profiles, also time series “derived” data (e.g. rollups, aggregations).
- 29. Five “V”s IoT data Velocity Torrent of small writes (sensors). Reads – millions of low-latency queries, user and device profiles, range queries for TS data (slices). Stream of updates (profiles) - beware of conflicts. Variety Sensors data (time series), users and devices profiles, also time series “derived” data (e.g. rollups, aggregations). Volume Starts small, grows quickly, keeps coming 24x7x365 (nights, weekends and holidays). Spikes up on new model launches or successful marketing campaign. But can slow down, but will keep growing. Efficient data retention policy is critical to prevent overflows.
- 30. Five “V”s IoT data Velocity Torrent of small writes (sensors). Reads – millions of low-latency queries, user and device profiles, range queries for TS data (slices). Stream of updates (profiles) - beware of conflicts. Variety Sensors data (time series), users and devices profiles, also time series “derived” data (e.g. rollups, aggregations). Volume Starts small, grows quickly, keeps coming 24x7x365 (nights, weekends and holidays). Spikes up on new model launches or successful marketing campaign. But can slow down, but will keep growing. Efficient data retention policy is critical to prevent overflows. Veracity Generally trustworthy, but beware of “low cost” sensors with low accuracy. Sent over not- so-reliable transport - expect that some data will be corrupted or arrive late or might be lost. (Hopefully the devices were not hijacked or impersonated by hackers)
- 31. Five “V”s IoT data Velocity Torrent of small writes (sensors). Reads – millions of low-latency queries, user and device profiles, range queries for TS data (slices). Stream of updates (profiles) - beware of conflicts. Variety Sensors data (time series), users and devices profiles, also time series “derived” data (e.g. rollups, aggregations). Volume Starts small, grows quickly, keeps coming 24x7x365 (nights, weekends and holidays). Spikes up on new model launches or successful marketing campaign. But can slow down, but will keep growing. Efficient data retention policy is critical to prevent overflows. Veracity Generally trustworthy, but beware of “low cost” sensors with low accuracy. Sent over not- so-reliable transport - expect that some data will be corrupted or arrive late or might be lost. (Hopefully the devices were not hijacked or impersonated by hackers) Value Profiles and summaries are much more valuable than raw data samples. The value of “raw” time series quickly goes down was processed and clock advances. Aggregated (”derived”) data are more valuable than raw data. Exceptions: financial transactions, life support, nuclear plants, oil rigs, …
- 32. Five “V”s IoT data Velocity Torrent of small writes (sensors). Reads – millions of low-latency queries, user and device profiles, range queries for TS data (slices). Stream of updates (profiles) - beware of conflicts. Variety Sensors data (time series), users and devices profiles, also time series “derived” data (e.g. rollups, aggregations). Volume Starts small, grows quickly, keeps coming 24x7x365 (nights, weekends and holidays). Spikes up on new model launches or successful marketing campaign. But can slow down, but will keep growing. Efficient data retention policy is critical to prevent overflows. Veracity Generally trustworthy, but beware of “low cost” sensors with low accuracy. Sent over not- so-reliable transport - expect that some data will be corrupted or arrive late or might be lost. (Hopefully the devices were not hijacked or impersonated by hackers) Value Profiles and summaries are much more valuable than raw data samples. The value of “raw” time series quickly goes down was processed and clock advances. Aggregated (”derived”) data are more valuable than raw data. Exceptions: financial transactions, life support, nuclear plants, oil rigs, … Complexity Poly-structured using simple schemas and simple relations (usually implicit). Some data is treated as unstructured (”opaque”) for speed or flexibility.
- 33. What architecture would work ?
- 34. Architectural Blueprints • Lambda Architecture by Nathan Marz (ex-Twitter) • Kappa Architecture by Jay Kreps (Confluent) • Zeta Architecture by Jim Scott (MapR) • … and their variants Lambda Kappa Zeta
- 35. Data Processing Framework for IoT • Uses “Best of breed” OSS technologies • Combines two paradigms • “Speed Layer” – pipeline for Stream Processing for “Data in Motion” • “Serving Layer” – analytics for “Data in Motion” and “Data at Rest” • Every component is “Distributed by Design” • Collection Layer • Message Queue • Stream Processing • Data Storage (Database, Object System, Data Warehouse) • Query and Analytics Engines
- 36. Data Access Patterns Category Description R:W % Metadata & Profiles Devices & Users Many low latency small reads - all over the dataset. Occasional updates – possibly by different “actors” (web, device, app), conflicts need to be prevented or resolved. Fewer creates and deletes. 90:10 Time Series
- 37. Data Access Patterns Category Description R:W % Metadata & Profiles Devices & Users Many low latency small reads - all over the dataset. Occasional updates – possibly by different “actors” (web, device, app), conflicts need to be prevented or resolved. Fewer creates and deletes. 90:10 Time Series Ingested (“Raw”) Very high throughout of relatively small writes. Most reads are over recent time range “slice”. Updates are rare (corrections). This category is a biggest part of the IoT application dataset. 10:90
- 38. Data Access Patterns Category Description R:W % Metadata & Profiles Devices & Users Many low latency small reads - all over the dataset. Occasional updates – possibly by different “actors” (web, device, app), conflicts need to be prevented or resolved. Fewer creates and deletes. 90:10 Time Series Ingested (“Raw”) Very high throughout of relatively small writes. Most reads are over recent time range “slice”. Updates are rare (corrections). This category is a biggest part of the IoT application dataset. 10:90 Aggregated (“Derived”) Mostly reads – users, platform services, reports. Writes are periodical on each time interval or from batch jobs. 80:20
- 39. Data store for IoT – “Wish list” • Ingested (Raw) Time Series • Very high write throughput • Fast slice (time range) reads • Aggregated (Derived) Time Series • Auto-distributed + slice locality • SQL-like queries • Aggregations • Bulk queries (analytics) • Secondary Indexes (Tags) • Efficient Storage • Auto Data Retention (TTL) • Build-in anti entropy • Compression • Hot Backups • Profiles and Metadata • Many concurrent reads with low latency • Reliable writes (ACID or conflict resolution) • Unstructured or partially structured • Secondary Indexes + Text Search • Scalability and Availability • Distributed architecture, no SPoF • Linearly scalable - up and down • Operational simplicity • Master-less architecture • Automatic rebalancing • Metrics, logs, events • Rolling upgrades
- 40. What DB type is a good fit for TS use cases?
- 41. Database Type For IoT or Time Series Relational Key Value Document Wide Column Graph MySQL Riak KV MongoDB Cassandra Neo4J PostgreSQL DynamoDB CouchBase HBase Titan Oracle Voldemort RethinkDB Accumulo Infinite Graph There is a need for a new type of NoSQL database – Time Series None of existing DB types was designed to handle time series data • Wide column DBs have high write throughput, but reads and updates are not their strength • Key Value and Document DBs handle metadata well, but struggle with heavy writes and time-slicing reads • Relational - good with metadata (unless number of updates is high), but a bad choice for TS data • Graph DB – not a good choice for either time series or metadata, can be added later on
- 42. Database Type For IoT or Time Series Relational Key Value Document Wide Column Graph MySQL Riak KV MongoDB Cassandra Neo4J PostgreSQL DynamoDB CouchBase HBase Titan Oracle Voldemort RethinkDB Accumulo Infinite Graph Time Series InfluxDB Riak TS Blueflood KairosDB Prometeus Druid OpenTSDB Dalmatiner Graphite
- 43. Iot Sensors Data – Hot to Cold
- 44. SENSORS DATA – HOT N’ COLD Temp Purpose Description Immutable? Boiling Hot App usage Last known value(s) and/or for last N minutes, useful for immediate responses, very frequently accessed No Hot Operational dataset Last 24 hours to several days or weeks (rarely months), frequently accessed, dashboards and online analytics Almost* Warm Historical data Older data, less frequently accessed, used mostly for offline analytics and historical analysis Yes Cold Archives Used only in rare situations, kept in long term storage for regulatory or unpredicted purposes Yes
- 45. STORAGE TIERS – FROM HOT TO COLD RAM → Database (TSDB) → Object Storage → Archive Data Lake Temp Purpose Storage Products Immutable? Boiling Hot App usage Internal app cache, Redis or Memcached No Hot Operational dataset NoSQL Database (preferably Time Series DB) Riak TS, OpenTSDB, KairosDB, Cassandra, HBase Almost* Warm Historical data Object storage – HDFS (Hadoop), Ceph, Minio, Riak S2 or AWS S3 Yes Cold Archives Various Yes
- 46. STORAGE TIERS – REALITY CHECK RAM → Database (TSDB) → Object Storage → Archive Elastic Cache (Redis) → Database (Postgres, DynamoDB) → AWS S3 → Glacier Data Lake Temp AWS Service Storage price, GB per month Boiling Hot Elastic Cache (Redis) $15-45 Hot DynamoDB RDS (Postgres) $ 0.25-0.35 (SSD) from $0.1 (Magnetic) Warm Simple Storage Service (S3) $0.024 to $0.030 Cold Glacier $0.007
- 47. OSS technologies for scalable IoT apps Component Open Source Technologies Load balancer Ngnix, HA Proxy Ingestion Kafka, RabbitMQ, ZeroMQ, Flume Stream Computing Spark Streaming, Apache Flink, Kafka Streams, Samza Time Series Store InfluxDB, KairosDB, Riak, Cassandra, OpenTSDB Profiles Store CouchBase, Riak, MySQL, Postgres, MongoDB Search Solr, Elastic Search Object Storage HDFS (Hadoop), Minio, Riak S2, Ceph Analytics Framework Apache Spark, MapReduce, Hive SQL Query Engine Spark SQL, Presto, Impala, Drill Cluster Manager Mesosphere DC/OS or Mesos, Kubernetes, Docker Swarm
- 48. Checklist for IoT technology stack ❑Is it vendor lock-in or open source software? Are there open APIs? ❑Can it be deployed in cloud? At the edge? In a data center? Using hybrid approach? ❑Can it be used it for free or low cost (no big upfront investment)? ❑Can you develop your app on your laptop? How many “moving parts”? ❑Are the components pre-integrated or can be easily integrated together? ❑Can you easily scale each component in this architecture by 10x? 20x? 50x? ❑Is there a roadmap, actively worked on, which is aligned with your vision? ❑Is there a company behind the technology to provide 24x7 support when needed?
- 49. Come to Basho booth to learn about • Riak TS (Time Series) - highly scalable NoSQL database for IoT and Time Series … and more • Riak Spark Connector for Apache Spark • Riak Integrations with Redis and Kafka • Riak Mesos Framework (RMF) for DC/OS
- 50. QUESTIONS?