Transis Dynamic Voting for Consistent Primary Components PODC

Transis Dynamic Voting for Consistent Primary Components PODC 1997 talk slides Esti Yeger Lotem, Idit Keidar and Danny Dolev The Hebrew University http: //www. cs. huji. ac. il/~transis

Transis Primary Components • Allows one subset of the processes to function when failures occur: – Database applications – Group communication systems (e. g. , ISIS) • Often based on majority (quorum) – Difficult to adapt to dynamic changes in the set of participants – Problematic in unreliable networks

Transis Dynamic Voting • Defines quorums adaptively: majority of the previous quorum, e. g. , • {a, b, c, d, e} • {a, b, c} • {a, b} • Naturally adapts to dynamic changes in the set of participants • In unreliable networks, proven to lead to better performance: • simulations, empirical tests, stochastic analysis

Transis Dynamic Linear Voting • Breaks ties between groups of equal size • Uses a linear order, L, on all potential processes in the system • Sub_Quorum(S, T) if: – T contains a majority of S, or – T contains half the members of S including the member, p, of S with the highest L(p)

Our Dynamic Voting Protocol Transis for primary components • Consistent • Allows processes to join and leave on the fly • Efficient – Low communication overhead • Simple to implement • Robust – Processes and links may fail

The Challenge: Transis Coping with failures that occur in the course of the protocol • {a, b, c} attempt to form a quorum • a and b succeed • c detaches, unaware of the attempt

The Challenge (Cont’d) Transis • {a, b} form a quorum – majority of {a, b, c} • Concurrently {c, d, e} form a quorum – majority of {a, b, c, d, e} ðInconsistency!

Transis Other Protocols • Inconsistent • Two-phase Commit – Limits the availability • ISIS - Cold start when primary is lost • Phoenix - Three phase consensus protocol – High communication overhead

Transis Our Solution: A Second Level of Knowledge • If a and b succeed in forming {a, b, c} then c is aware of the attempt. • For c, {a, b, c} is ambiguous: {a, b, c} may or may not have been formed. – Processes record ambiguous attempts • In our example, c records both: {a, b, c, d, e} and {a, b, c} – Requires a majority of both ð c will refuse to form {c, d, e}

A Session of the Protocol Transis General Scheme • Invoked Upon Membership Changes 1 Exchange information 2 If Sub_Quorum of the last primary and of all ambiguous attempts, Attempt: – Record the attempt as ambiguous 3 If all attempted, Form: – Become the primary in the system Delete all ambiguous attempts

Transis Storing all the Ambiguous Attempts is not Feasible • When failures cascade, the number of ambiguous attempts may be exponential: {a, b, c, d, e}, {a, b, c}, {a, b, d}, {a, b, e}, {a, b, c, f, g}, {a, b, c, f}, {a, b, d, g}, … • Ambiguous attempts Þ constraints • We use a “garbage collection” mechanism to store only a linear number of attempts

Transis Our “Garbage Collection” Mechanism • Resolution rules: – If the attempt was formed by some member, adopt it as your primary. – If the attempt was not formed by any of the members, delete it. • Learning rules: – p learns the status of q w. r. t. attempt A 1 during a later attempt A 2 • Linear

Why is this linear? Transis (In the number of processes) • If p and q participate in two attempts, A 1 and A 2 then: p learned whether q formed A 1 before attempting to form A 2 • Once p learns about A 1 from all its members, p can resolve A 1 ð In each recorded attempt there is a member that does not appear in later attempts

Transis Criticism of Dynamic Voting • Quorums can become very small (even one process) – Failure of a single process may cause the rest of the system to block • Desirable solution: – Set a threshold, Min_Quorum, on quorum size – (N - Min_Quorum) process are always a quorum – Min_Quorum reflects the tradeoff between “static” and “dynamic”

Transis The Challenge • What happens if N (the number of processes) changes on the fly? – N - Min_Quorum changes – The “truth value” of Sub_Quorum changes (no longer a predicate) • Asynchronous distributed system – Different processes may know of different values of N

Adding New Processes Carefully Transis - two steps • W - The set of participating processes – may “vote” for quorums • A - Candidates to be added to W – do not “vote” for quorums – “vote against” large quorums • New processes are added: – to A in the attempt step, – to W in the form step

Transis The Min_Quorum Requirement • Every quorum must contain more than Min_Quorum members of W – at least Min_Quorum “vote for” • Every group that contains all but Min_Quorum members of W and A is a quorum (regardless of past quorums) – At most Min_Quorum “vote against”

Transis Conclusions • Consistently maintaining a primary component using dynamic voting • More available than other protocols • Simple and efficient • No need for cold start • New mechanism: always allowing large groups to be quorums where processes can join on the fly