Cohesity, Inc.

Vinay Reddy

Mohit Aron

Vipin Gupta

Markose Thomas

Sashikanth Madduri

Techniques are disclosed for allocating shared resources to nodes in a distributed computing network system. Nodes request a lock for each instance of a computing resource (e.g., a virtual IP address associated with a service provided by the distributed computing network system) from a distributed lock manager. The distributed lock manager maintains a queue of requests for each instance of the shared resource. Upon receiving a lock from the distributed lock manager, the receiving node performs a fairness allocation protocol to determine whether to accept the lock. If so determined, the shared computing resources associated with the lock is configured.

Distributed fair allocation of shared resources to constituents of a cluster

Backup data associated with a primary storage system is received. At least a view of the received backup data associated with the primary storage system is generated. The view of the of the received backup data associated with the primary storage system includes a tree data structure comprising a root node, a first plurality of child nodes, and a first plurality of leaf nodes. At least some of the first plurality of leaf nodes include corresponding pointers to a corresponding binary large object. A read request for data exposed by the view of the received backup data associated with the primary storage system is received from an external system. In response to receiving the read request, the view of the received backup data associated with the primary storage system is exposed to a requesting system utilizing a protocol associated with the external system.

Making more active use of a secondary storage system

Techniques are disclosed which allow a secondary storage system to provide data to non-production workloads in conjunction with performing data backup and protection tasks. As disclosed, a secondary storage system exposes backup data stored by the secondary storage system to other workloads, such as test and development applications, data analytics, etc. These non-production workloads can run at the same time the secondary storage system provides backup services to a primary storage system. This consolidation eliminates the need for an enterprise to deploy separate storage clusters for analytics, test and development applications, etc. and eliminates unnecessary copies of data.

Techniques are disclosed for managing a high performance, fault-tolerant, strongly consistent, distributed key-value store system. The key-value store may store information, such as metadata for a distributed file system. Fault-tolerance means that the distributed key-value store continues to provide access to values in the key-value store in spite of a certain number of node failures. To provide this capability, the key-value store may store copies of (key, value) pair on N+1 nodes in order to provide fault tolerance for the failure of up to N nodes. In addition, metadata describing which nodes store a given value is stored on 2N+1 nodes and the distributed key-value store is sized such that there are 3N+1 nodes in a cluster. Doing so allows the key, value store to tolerate a failure of N nodes, while still maintaining a consistent and available key-value store.

Performing file system operations in a distributed key-value store

A lock associated with a computing resource is received. It is determined whether to accept the lock based on a fairness allocation protocol. The lock is accepted based on a determination of the fairness allocation protocol.

Systems, methods, and computer program products for managing a consensus group in a distributed computing cluster, by determining that an instance of an authority module executing on a first node, of a consensus group of nodes in the distributed computing cluster, has failed; and adding, by an instance of the authority module on a second node of the consensus group, a new node to the consensus group to replace the first node. The new node is a node in the computing cluster that was not a member of the consensus group at the time the instance of the authority module executing on the first node is determined to have failed.

Growing community of inventors

San Jose, CA, United States of America

Vinay Reddy