Amazon Technologies, Inc.

Xianglong Huang

Wei Xiao

David Alan Lutz

Somasundaram Perianayagam

Yijun Lu

Jiandan Zheng

Maximiliano Maccanti

Stuart Henry Seelye Marshall

Rande Anthony Blackman

Kiran Kumar Muniswamy Reddy

Pejus Manoj Das

Stefano Stefani

Timothy Andrew Rath

Kiran-Kumar Muniswamy-Reddy

Rama Krishna Sandeep Pokkunuri

Jai Vasanth

Adam Douglas Morley

Barry Bailey Hunter, Jr

Yaser Al-Onaizan

Clarence Wing Yin Ng

Jian Wang

Lokendra Singh Panwar

Marcello Federico

Georgiana Dinu

Prashant Mathur

Nirmal Mehta

Aakash Upadhyay

Eric J Nowell

Haiyu Yao

Anna Currey

A translation system receives examples of translations between a first language and a second language. In response to receiving request to translate a source text from the first language to the second language, the system ranks the examples based on the example's applicability to one or more portions of the source text. The system performs additional training of a neural network that was pre-trained to translate from the first language to the second language, where the additional training is based on one or more top-ranking examples. The system translates the source text to the second language using the additionally trained neural network.

Machine translation with adapted neural networks

One or more table partitions may communicate with an index partition that may be a master of a replication group. A communications channel may exist between table partitions and the index partition. Upon splitting the index partition, communications between the table partitions and the index partition may be suspended. Upon completion of the split, communications may be reestablished between the table partitions and a partition, of the replication group of index partitions, designated to be a master following the split. Messages accumulated by the table partitions during the split may be sent to the index partition upon reestablishing communications.

Table and index communications channels

A distributed database management system may comprise a plurality of computing nodes. A request to update an item maintained by the system may be acknowledged as durable and committed once an entry corresponding to the request has been written to a log file and quorum among the computing nodes has been achieved. Improved consistency may be achieved by maintaining snapshots of committed item states within queryable in-memory snapshot data structures. Range queries may be performed by merging a secondary index with the snapshots and applying filters. Projections may be completed by retrieving additional data from an item collection maintain on one or more storage devices.

Consistent query of local indexes

A data storage system may implement implicit checkpoint for generating a secondary of a table. Indexing updates may be replicated and maintained across a replica group storing a table for a data store. Upon detection of a restart event for generating a secondary index, a replica in the replica group may evaluate the indexing updates to determine an index creation restart point according to an order for indexing the table. The generation of the secondary index may be resumed at the index creation restart point. In this way, secondary index generation may continue whether or not a previously indexing replica in the replica group, such as a master replica, is available to continue generating the secondary index.

Implicit checkpoint for generating a secondary index of a table

A data storage system may implement rate-limiting secondary index creation for an online table. A secondary index may be generated for a table stored in a data store. The table may be incrementally indexed, maintaining the updates determined according to indexing different portions of the table in a queue of pending updates that are subsequently applied at the secondary index. Prior to indexing a portion of the table, an evaluation of a current number of pending updates in the queue of pending updates may be performed with respect to a throttle threshold. If the current number of pending updates exceeds the throttle threshold, then indexing the portion of the table may be throttled. Received updates to previously indexed portions of the table, may be applied to the table and placed in the queue of pending updates without an evaluation of the current number of pending requests.

Rate-limiting secondary index creation for an online table

One or more table partitions may communicate with one or more index partitions. A communications channel may be dedicated to each combination of table partition and index partition. The communications channel may act as a queue of messages corresponding to updates to data maintained in a table partition that have yet to be processed by a corresponding index partition. Upon splitting a table partition, communications channels may be duplicated including unprocessed messages. Connections between new table partitions and an index may be finalized upon the index's receipt of a message indicative of the new table partition.

A data storage system may implement locking item ranges for creating a secondary index of an online table. A secondary index may be generated for a table of items stored in a non-relational data store. Different ranges of items in the data store may be locked while a corresponding portion of the secondary index is generated. Upon generating the corresponding portion of the secondary index, a range of items may be unlocked. While generating the secondary index, the table may be made available for servicing access requests. For a request to update the table received during the generation of the secondary index, a determination may be made as to whether the update is included within a locked range of the table. If locked, the request may be delayed until the range is unlocked.

Locking item ranges for creating a secondary index from an online table

A system that implements a data storage service may store data in multiple replicated partitions on respective computing nodes on behalf of clients. A storage node may, based on the amount of provisioned resources on a given storage device or logical volume, identify candidate partition management operations to be performed, and may send information about the operations to a central partition management scheduler. The scheduler may apply a global prioritization scheme to determine an order in which to perform the candidate operations. The order may be based on whether the operations include partition splits or partition moves, whether they aim to reduce provisioned storage capacity or reduce throughput capacity on a storage device or logical volume, whether they conflict with each other, whether the total number of partitions (or replicas thereof) involved in partition management at any given time exceeds a pre-determined limit, or whether they were requested by clients.

Scheduling of splits and moves of database partitions

An automated system may be employed to perform detection, analysis and recovery from faults occurring in a distributed computing system. Faults may be recorded in a metadata store for verification and analysis by an automated fault management process. Diagnostic procedures may confirm detected faults. The automated fault management process may perform recovery workflows involving operations such as rebooting faulting devices and excommunicating unrecoverable computing nodes from affected clusters.

Distributed computing fault management

A distributed database management system comprising a plurality of computing nodes may distribute data evenly across all nodes. A definition of a primary key that divides the primary key into at least a first key portion and a second key portion may be utilized to locate items related by a first key portion to a specific computing node. Application-consistent queries, local transactions and pivoting operations may be performed on items related by a first key portion.

Item collections

Growing community of inventors

Bellevue, WA, United States of America

Xianglong Huang