Two-phase commit optimizations and tradeoffs in the commercial environment

Abstract

A n atomic commit protocol can ensure that all participants rn a distributed transactron reach consistent states, whether or not system or network farlures occur. One widely used protocol is the two-phase commit (2PC) protocol, whrch has long appeared in the literature.Much of the lrterature focuses on improving performance in failure cases by providing a non blocking 2PC that streamlines recovery processrng at the expense of extra processrng in the normal case. We focus on improvrng performance in the normal case based on two assumptrons: first that networks and systems are becoming increasingly reliable, and second that the need to support high-volume transactions requires a streamlined protocol for the normal case.In this paper, various optimizations are analyzed in terms of reliability. savings in log writes and network traffic, and reductron in resource lock trme.Its unique contributrons Include the description of some optimizations not descrrbed elsewhere and a systematic comparison of the optimrzatrons and the environments where they cause the most benefit.

Distributed real time database systems: background and literature review

“…It ensures that sufficient information is force-written on the stable storage to reach a consistent global decision about the transaction [9,102,131]. A number of 2PC variants [84,98,121] have been proposed and can be classified into following four groups [86].…”

OSI distributed transaction processing commitment optimizations

“…By giving up explicit PREPARE-messages, the optimization 'Unsolicited Ready' [8] (named 'Unprompted Ready' in the OSI TP protocol) is achieved. From the point of view of a correct application, a READY-message will never be received unexpectedly.…”

“…By this, the condition for sending a READY-message is generalised to: "A node having received READY-messages from all neighbours but one writes a log ready record, enters the READY-state and issues a READY-message to the one neighbour, from which no READY-message has been received." This generalised rule is the basis for the well known optimization called 'Last Subordinate Optimization' (LSO) [8].…”

“…There are very different reasons for transferring the role of the commitment coordinator from the root node of the transaction tree to another node and therefore to liberalise the flow of READY-messages: (1) reduction of runtime of a transaction [8], [9], and (2) allowance for different degrees of availability of the participating systems (see section 4.1.2 for a detailed discussion).…”

Data replication in Mariposa

“…Classic distributed data managers require that copies be kept fully consistent. Because of this, considerable eort has gone into improving the basic techniques for ensuring this kind of consistency, such a s t w o-phase commit [ELAB85,SAMA93]. However, because the expense of synchronizing updates remains relatively high, work has also been done in constructing weak consistency models.…”