Dropbox, Inc.

Amazon Technologies, Inc.

Bogdan Munteanu

Mehant Baid

Daniel K Tahara

Jonathan Robert Collins

Mihnea Cezar Giurgea

Divya Malika Patruni

Neha Sherpa

Computer-implemented techniques for replication-lag constrained deletion of data in a distributed data storage system. In some aspects, the techniques improve the operation of a computing system by preventing too high of a delete rate that causes severe replication lag while at the same time increasing and decreasing the delete rate over time to a maximum allowable delete rate constrained by measured replication lag in terms of both local replication lag and geographic replication lag. In one implementation, the delete rate is adjusted by increasing or decreasing a pause interval that determines how long a database data deletion process pauses between submitting database deletion commands to a database server.

Replication lag-constrained deletion of data in a large-scale distributed data storage system

Computer-implemented techniques for automatically coordinating application schema changes in a distributed data storage system. The techniques improve the operation of a computing system by automating the deployment of application schemas changes to a distributed data storage system in a manner that does not introduce orphaned data inconsistencies or integrity inconsistencies to the application data stored in the distributed data storage system. The techniques include a schema change deployment protocol that ensures that all of the database servers of the distributed data storage system are using no more than two consecutive versions of the application schema at any given time.

Automatically coordinating application schema changes in a distributed data storage system

A method for performing auto-repair of a data volume may include determining a number of unavailable shards from a plurality of shards associated with a data volume. The plurality of shards may be stored in a plurality of data centers. The data volume may be erasure-coded with a total of N shards and can be reconstructed using at least K available shards out of the N shards. The unavailable shards may be filtered to exclude shards that have been unavailable for less than a threshold amount of time. For each of the plurality of data centers, a total number of unavailable shards may be determined if the data center is disabled. The total number of unavailable shards may be based at least in part on a number of shards associated with the data volume and stored at the data center, and a number of the filtered unavailable shards stored at the remaining plurality of data centers.

Data volume auto-repair based on volume degradation level

Growing community of inventors

San Mateo, CA, United States of America

Bogdan Munteanu