Near Real-Time Ingestion
Ingesting data from external sources like (event logs, databases, external sources) into a Hadoop Data Lake is a well known problem. In most (if not all) Hadoop deployments, it is unfortunately solved in a piecemeal fashion, using a medley of ingestion tools, even though this data is arguably the most valuable for the entire organization.
For RDBMS ingestion, Hoodie provides faster loads via Upserts, as opposed costly & inefficient bulk loads. For e.g, you can read the MySQL BIN log or Sqoop Incremental Import and apply them to an equivalent Hoodie table on HDFS. This would be much faster/efficient than a bulk merge job or complicated handcrafted merge workflows
For NoSQL datastores like Cassandra / Voldemort / HBase, even moderately big installations store billions of rows. It goes without saying that full bulk loads are simply infeasible and more efficient approaches are needed if ingestion is to keep up with the typically high update volumes.
Even for immutable data sources like Kafka , Hoodie helps enforces a minimum file size on HDFS, which improves NameNode health by solving one of the age old problems in Hadoop land in a holistic way. This is all the more important for event streams, since typically its higher volume (eg: click streams) and if not managed well, can cause serious damage to your Hadoop cluster.
Across all sources, Hoodie adds the much needed ability to atomically publish new data to consumers via notion of commits, shielding them from partial ingestion failures
Near Real-time Analytics
Typically, real-time datamarts are powered by specialized analytical stores such as Druid or Memsql or even OpenTSDB . This is absolutely perfect for lower scale (relative to Hadoop installations like this) data, that needs sub-second query responses such as system monitoring or interactive real-time analysis. But, typically these systems end up getting abused for less interactive queries also since data on Hadoop is intolerably stale. This leads to under utilization & wasteful hardware/license costs.
On the other hand, interactive SQL solutions on Hadoop such as Presto & SparkSQL excel in queries that finish within few seconds. By bringing data freshness to a few minutes, Hoodie can provide a much efficient alternative, as well unlock real-time analytics on several magnitudes larger datasets stored in HDFS. Also, Hoodie has no external dependencies (like a dedicated HBase cluster, purely used for real-time analytics) and thus enables faster analytics on much fresher analytics, without increasing the operational overhead.
Incremental Processing Pipelines
One fundamental ability Hadoop provides is to build a chain of datasets derived from each other via DAGs expressed as workflows.
Workflows often depend on new data being output by multiple upstream workflows and traditionally, availability of new data is indicated by a new HDFS Folder/Hive Partition.
Let’s take a concrete example to illustrate this. An upstream workflow
U can create a Hive partition for every hour, with data for that hour (event_time) at the end of each hour (processing_time), providing effective freshness of 1 hour.
Then, a downstream workflow
D, kicks off immediately after
U finishes, and does its own processing for the next hour, increasing the effective latency to 2 hours.
The above paradigm simply ignores late arriving data i.e when
event_time drift apart.
Unfortunately, in today’s post-mobile & pre-IoT world, late data from intermittently connected mobile devices & sensors are the norm, not an anomaly.
In such cases, the only remedy to guarantee correctness is to reprocess the last few hours worth of data,
over and over again each hour, which can significantly hurt the efficiency across the entire ecosystem. For e.g; imagine reprocessing TBs worth of data every hour across hundreds of workflows.
Hoodie comes to the rescue again, by providing a way to consume new data (including late data) from an upsteam Hoodie dataset
HU at a record granularity (not folders/partitions),
apply the processing logic, and efficiently update/reconcile late data with a downstream Hoodie dataset
HD can be continuously scheduled at a much more frequent schedule
like 15 mins, and providing an end-end latency of 30 mins at
Data Dispersal From Hadoop
A popular use-case for Hadoop, is to crunch data and then disperse it back to an online serving store, to be used by an application.
For e.g, a Spark Pipeline can determine hard braking events on Hadoop and load them into a serving store like ElasticSearch, to be used by the Uber application to increase safe driving. Typical architectures for this employ a
queue between Hadoop and serving store, to prevent overwhelming the target serving store.
A popular choice for this queue is Kafka and this model often results in redundant storage of same data on HDFS (for offline analysis on computed results) and Kafka (for dispersal)
Once again Hoodie can efficiently solve this problem, by having the Spark Pipeline upsert output from each run into a Hoodie dataset, which can then be incrementally tailed (just like a Kafka topic) for new data & written into the serving store.