Computing and tracking client staleness using transaction responses

Abstract

Distributed systems that maintain data items or state on multiple server nodes (e.g., a storage system or lock service) may receive transaction requests from clients that include read requests and/or write requests. If a transaction request includes a read request, a server node may determine a node staleness value for the requested data from the server's perspective, and may return a transaction response to the client that includes the requested data and the node staleness value. The client may compute a client staleness value for the requested data dependent on the node staleness value in the transaction response, rather than on a separate heartbeat process. If the transaction also includes a write request, the server may return the transaction response for the read request before the write commits. In this case, the client may not use the requested data until it receives another response indicating that the write committed.