Reliable Distributed Systems

Membership

Agreement on Membership Recall our approach:

Detecting failure is a lost cause. Too many things can mimic failure To be accurate would end up waiting for a process

to recover Substitute agreement on membership

Now we can drop a process because it isn’t fast enough

This can seem “arbitrary”, e.g. A kills B… GMS implements this service for

everyone else

Architecture

Membership Agreement, “join/leave” and “P seems to be unresponsive”

2PC-like protocols use membership changes instead of failure notification

Applications use replicated data for Applications use replicated data for high availabilityhigh availability

Architecture

GMS

A

B

C

D

join leave

join

A seems to have failed

{A}

{A,B,D}

{A,D}

{A,D,C}

{D,C}

X Y Z

Application processes

GMS processes

membership views

Contrast dynamic with static model Static model: fixed set of processes “tied” to

resources Processes may be unreachable (while failed or

partitioned away) but later recover Think: “cluster of PCs”

Dynamic model: changing set of processes launched while system runs, some fail/terminate

Failed processes never recover (partitioned process may reconnect, but uses a new pid)

And can still own a physical resource, allowing us to emulate a static model

Consistency options Could require that system always be

consistent with actions taken at a process even if that process fails immediately after taking the action This property is needed in systems that take

external actions, like advising an air traffic controller

May not be needed in high availability systems Alternative is to require that operational

part of system remain continuously self-consistent

Obstacles to progress Fischer, Lynch and Patterson result:

proof that agreement protocols cannot be both externally consistent and live in asynchronous environments

Suggests that choice between internal consistency and external consistency is a fundamental one!

Can show that this result also applies to dynamic membership problems

Usual response to FLP: Chandra/Toueg Consider system as having a failure

detector that provides input to the basic system itself

Agreement protocols within system are considered safe and live if they satisfy their properties and are live when the failure detector is live

Babaoglu: expresses similar result in terms of reachability of processes: protocols are live during periods of reachability

Towards an Alternative In this lecture, focus on systems

with self-defined membership Idea is that if p can’t talk to q it will

initiate a membership change that removes q from p’s system “membership view”

Illustrated on next slide

Commit protocol from when we discussed transactions

ok to commit?

okdecision unknown!

vote unknown!ok

Suppose this is a partitioning failure

ok to commit?

okdecision unknown!

vote unknown!ok

Do these processes actually need to be consistent with the others?

Primary partition concept Idea is to identify notion of “the system”

with a unique component of the partitioned system

Call this distinguished component the “primary” partition of the system as a whole. Primary partition can speak with authority

for the system as a whole Non-primary partitions have weaker

consistency guarantees and limited ability to initiate new actions

Ricciardi: Group Membership Protocol For use in a group membership service

(usually just a few processes that run on behalf of whole system)

Tracks own membership; own members use this to maintain membership list for the whole system

All user’s of the service see subsequences of a single system-wide group membership history

GMS also tracks the primary partition

GMP protocol itself Used only to track membership of the “core”

GMS Designates one GMS member as the

coordinator Switches between 2PC and 3PC

2PC if the coordinator didn’t fail and other members failed or are joining

3PC if the coordinator failed and some other member is taking over as new coordinator

Question: how to avoid “logical partitioning”?

GMS majority requirement To move from system “view” i to view

i+1, GMS requires explicit acknowledgement by a majority of the processes in view i

Can’t get a majority: causes GMS to lose its primaryness information

Dahlia Malkhi has extended GMP to support partitioning and remerging; similar idea used by Yair Amir and others in Totem system

GMS in Action

p0

p1

...

p5

p0 is the initial coordinator. p1 and p2 join, then p3...p5 join. But p0 fails during join protocol, and later so does p3. Notice use of majority consent to avoid partitioning!

GMS in Action

p0

p1

...

p5

2-phase commit… 3-phase… 2–phaseP0 is coordinator… P1 takes over… P1 is new coordinator

What if system has thousands of processes?

Idea is to build a GMS subsystem that runs on just a few nodes

GMS members track themselves Other processes ask to be admitted to

system or for faulty processes to be excluded

GMS treats overall system membership as a form of replicated data that it manages, reports to its “listeners”

Uses of membership?

If we rewire TCP and RPC to use membership changes as trigger for breaking connections, can eliminate split-brain problems! But nobody really does this Problem is that networks lack

standard GMS subsystems now! But we can still use it ourselves

Replicated data within groups A very general requirement:

Data actually managed by group Inputs and outputs, in a server replicated for

fault-tolerance Coordination and synchronization data

Will see how to solve this, and then will use solution to implement “process groups” which are subgroups of the overall system membership

Replicated data Assume that we have a (dynamically

defined) group of processes G and that its members manage a replicated data item

Goal: update by sending a multicast to G

Should be able to safely read any copy “locally”

Consider situation where members of G may fail or recover

Some Initial Assumptions For now, assume that we work directly on the

real network, not using Ricciardi’s GMS Later will need to put GMS in to solve a

problem this raises, but for now, the model will be the very simple one: processes that communicate using messages, asynchronous network, crash failures

We’ll also need our own implementation of TCP-style reliable point-to-point channels using GMS as input

Process group model Initially, we’ll assume we are simply

given the model Later will see that we can use reliable

multicast to implement the model First approximation: a process group is

defined by a series of “views” of its membership. All members see the same sequence of view changes. Failures, joins reported by changing membership

Process groups with joins, failures

crash

G0={p,q} G1={p,q,r,s} G2={q,r,s} G3={q,r,s,t}

p

q

r

s

tr, s request to join

r,s added; state xfer

t added, state xfer

t requests to join

p fails

State transfer

Method for passing information about state of a group to a joining member

Looks instantaneous, at time the member is added to the view

Outline of treatment First, look at reliability and failure

atomicity Next, look at options for “ordering” in

group multicast Next, discuss implementation of the

group view mechanisms themselves Finally, return to state transfer Outcome: process groups, group

communication, state transfer, and fault-tolerance properties

Atomic delivery Atomic or failure atomic delivery

If any process receives the message and remains operational, all operational destinations receive it

p

q

r

s

a

b

All processes that receive a subsequently fail.

failsfails

All processes receive b.

Additional properties A multicast is dynamically uniform

if: If any process delivers the multicast,

all group members that don’t fail will deliver it (even if the initial recipient fails immediately after delivery).

Otherwise we say that the multicast is “not uniform”

Uniform and non-uniform delivery

p

q

r

s

a

b

Uniform delivery of a and b

failsfails

p

q

r

s

a

b

Non-uniform delivery of a

failsfails

Stronger properties cost more

Weaker ordering guarantees are cheaper than stronger ones

Non-uniform delivery is cheap Dynamic uniformity is costly Dynamic membership is cheap Static membership is more costly

Conceptual cost graph

less ordered local total order global total order

non

-un

ifor

m, d

ynam

ic g

rou

p

u

nif

orm

sta

tic

grou

p

asynchronous and non-uniform “cbcast” to dynamically defined group

uniform and globally total “abcast” in a static group

cbcast in Horus: 85,000/second, 85us latency sender to dest

Total, safe abcast in Totem or Transis: 600/second, 750ms latency sender to dest

Implementing multicast primitives Initially assume a static process group Crash failures: permanent failures, a

process fails by crashing undetectably. No GMS (at first).

Unreliable communication: messages can be lost in the channels

... looks like the asynchronous model of FLP

Failures? Message loss: overcome with retransmission Process failures: assume they “crash” silently Network failures: also called “partitioning” Can’t distinguish between these cases!

p

q

timeout: q failed!

timeout: p failed!

network partitions

Multicast by “flooding” All recipients echo message to all other

recipients, O(n2) messages exchanged Reject duplicates on basis of message id

When can we garbage collect the id?

p

q

r

s

a failsfails

Multicast by “flooding” All recipients echo message to all other

recipients, O(n2) messages exchanged Reject duplicates on basis of message id

When can we garbage collect the id?

p

q

r

s

a

Multicast by “flooding” All recipients echo message to all other

recipients, O(n2) messages exchanged Reject duplicates on basis of message id

When can we garbage collect the id?

p

q

r

s

a failsfails

Multicast by “flooding” All recipients echo message to all other

recipients, O(n2) messages exchanged Reject duplicates on basis of message id

When can we garbage collect the id?

p

q

r

s

a failsfails

Garbage collection issue Must remember id as long as might still

see a duplicate copy If no process fails: garbage collect after

echoed by all destinations Very similar to 3PC protocol

... correctness of this protocol depends upon having an accurate way to detect failure! Return to this point in a few minutes.

“Lazy” flooding and garbage collection Idea is to delay “non urgent” messages Recipients delay the echo in hope that sender

will confirm successful delivery: O(n)

messages p

q

r

s

a

ack...

“Lazy” flooding Recipients delay the echo in hope that sender

will confirm successful delivery: O(n)

messages

p

q

r

s

a

ack... all got it...

“Lazy” flooding Recipients delay the echo in hope that

sender will confirm successful delivery: O(n) messages

Notice that garbage collection occurs in 3rd phase

p

q

r

s

a failsfails

ack... all got it... garbage collect

“Lazy” flooding, delayed phases “Background” acknowedgements (not shown) Piggyback 2nd, 3rd phase on other multicasts

p

q

r

s

m1

m1

“Lazy” flooding, delayed phases “Background” acknowedgements (not shown) Piggyback 2nd, 3rd phase on other multicasts

p

q

r

s

m1 m2

m1 m2, all got m

“Lazy” flooding, delayed phases “Background” acknowedgements (not shown) Piggyback 2nd, 3rd phase on other multicasts

p

q

r

s

m1 m2 m3

fails

m1 m2, all got m1 m3, gc m1

“Lazy” flooding, delayed phases

“Background” acknowedgements (not shown) Piggyback 2nd, 3rd phase on other multicasts

Reliable multicasts now look cheap!

p

q

r

s

m1 m2 m3 m4

m1 m2, all got m1 m3, gc m1 m4, gc m2

fails

Lazy scheme continued

If sender fails, recipients switch to flood-style algorithm... but now we have the same garbage

collection problem: if sender fails we may never be able to garbage collect the id!

Problem is caused by lack of failure detector

Garbage collection with inaccurate failure detections

... we lack an accurate way to detect failure If any does seem to fail, but is really still

operational and merely partitioned away, the connection might later be fixed.

That process might “wake up” and send a duplicate

Hence, if we are not sure a process has failed, can’t garbage collect our duplicate-supression data yet!

Exploiting a failure detector Suppose that we had a failstop environment Process group membership managed by

oracle, perhaps the GMS we saw earlier Failures reported as “new group views” All see the same sequence of views:

G = {p,q,r,s} {p,r,s} {r,s} Now can assume failures are accurately

detected

Now our lazy scheme works! Garbage collect when all non-faulty

processes are known to have received the message

Use process ranking to pick a new “coordinator” if the initial one fails

Cost only reaches n2 if many fail during protocol

Can delay 2nd, 3rd round if desired Also link GMS to point-to-point channel

implementation

Failure Detectors Needed as “input” to GMS. For now, just

assume we have one, perhaps Vogel’s investigator

In practice many systems use “timeout”, but timeout is not safe for our purposes

Feeding detections through group membership service converts inaccurate failure detections into what look like failstop failures for processes within the system

Cutting Channels to Failed Processes When a process is dropped from

the membership, break the connection to it

This will effectively eliminate the risk of “late” delivery of duplicate messages, etc.

Makes a partitioning failure look like a failstop failure.

Dynamic uniformity This property requires an extra phase of

communication Phase 1: distribute message Phase 2: can deliver if all non-faulty

processes received it in phase 1 Insight: no process delivers a message

until all have received it

Summary We know how to build a GMS that tracks

its own membership We know how to build an unordered

reliable multicast Actually, “sender-ordered” But from different senders, can delivery in

arbitrary orders And we know how to support various

forms of uniformity Next: multicast ordering