Cloudera, Inc.

Todd Lipcon

David Alves

Adar Lieber-Dembo

Eli Collins

Aaron T Myers

Columnar storage provides many performance and space saving benefits for analytic workloads, but previous mechanisms for handling single row update transactions in column stores suffer from poor performance. A columnar data layout facilitates both low-latency random access capabilities together with high-throughput analytical access capabilities, simplifying Hadoop architectures for use cases involving real-time data. In disclosed embodiments, systems may perform read and scan features through which users may have consistent reads of the columnar storage at any given point in time. These scans (e.g., snapshots) allow for analytics to be performed on the data, consistent backups of the data, for incremental backups, perform cross-cluster synchronization, and/or for offline audit analysis.

Snapshots of a column store

A first event occurs at a first computer at a first time, as measured by a local clock. A second event is initiated at a second computer by sending a message that includes the first time. The second event occurs at a second time, as measured by a local clock. Because of clock error, the first time is later than the second time. Based on the first time being later than the second time, an alternate second time, that is based on the first time, is used as the time of the second event. When a third system determines the order of the two events, the first time is obtained from the first computer, and the alternate second time is obtained from the second computer, and the order of the events is determined based on a comparison of the two times.

Ensuring properly ordered events in a distributed computing environment

Columnar storage provides many performance and space saving benefits for analytic workloads, but previous mechanisms for handling single row update transactions in column stores suffer from poor performance. A columnar data layout facilitates both low-latency random access capabilities together with high-throughput analytical access capabilities, simplifying Hadoop architectures for use cases involving real-time data. In disclosed embodiments, mutations within a single row are executed atomically across columns and do not necessarily include the entirety of a row. This allows for faster updates without the overhead of reading or rewriting larger columns.

Accumulating and flushing mutations in a column store

A compaction policy imposing soft limits to optimize system efficiency is used to select various rowsets on which to perform compaction, each rowset storing keys within an interval called a keyspace. For example, the disclosed compaction policy results in a decrease in a height of the tablet, removes overlapping rowsets, and creates smaller sized rowsets. The compaction policy is based on the linear relationship shared between the keyspace height and the cost associated with performing an operation (e.g., an insert operation) in that keyspace. Accordingly, various factors determining which rowsets are to be compacted, how large the compacted rowsets are to be made, and when to perform the compaction, are considered within the disclosed compaction policy. Furthermore, a system and method for performing compaction on the selected datasets in a log-structured database is also provided.

Database compaction in distributed data system

Systems and methods of a memory allocation buffer to reduce heap fragmentation. In one embodiment, the memory allocation buffer structures a memory arena dedicated to a target region that is one of a plurality of regions in a server in a database cluster such as an HBase cluster. The memory area has a chunk size (e.g., 2 MB) and an offset pointer. Data objects in write requests targeted to the region are received and inserted to the memory arena at a location specified by the offset pointer. When the memory arena is filled, a new one is allocated. When a MemStore of the target region is flushed, the entire memory arenas for the target region are freed up. This reduces heap fragmentation that is responsible for long and/or frequent garbage collection pauses.

Memory allocation buffer for reduction of heap fragmentation

A technique is described for merging multiple lists of ordinal elements such as keys into a sorted output. In an example embodiment, a merge window is defined, based on the bounds of the multiple lists of ordinal elements, that is representative of a portion of an overall element space associated with the multiple lists. Lists of elements to be sorted can be placed into one of at least two different heaps based on whether they overlap the merge window. For example, lists that overlap the merge window may be placed into an active or 'hot' heap, while lists that do not overlap the merge window may be placed into a separate inactive or 'cold' heap. A sorted output can then be generated by iteratively processing the active heap. As the processing of the active heap progresses, the merge window advances, and lists may move between the active and inactive heaps.

Growing community of inventors

San Francisco, CA, United States of America

Todd Lipcon