Cloud computing systems today, whether open-source or used inside companies, are built using a common set of core techniques, algorithms, and design philosophies – all centered around distributed systems. Learn about such fundamental distributed computing "concepts" for cloud computing. Some of these concepts include: clouds, MapReduce, key-value/NoSQL stores, classical distributed algorithms, widely-used distributed algorithms, scalability, trending areas, and much, much more! Know how these systems work from the inside out. Get your hands dirty using these concepts with provided homework exercises. In the programming assignments, implement some of these concepts in template code (programs) provided in the C++ programming language. Prior experience with C++ is required. The course also features interviews with leading researchers and managers, from both industry and academia. This course builds on the material covered in the Cloud Computing Concepts, Part 1 course.
2.2. Two-Phase Commit
Loading...
Del curso dictado por University of Illinois at Urbana-Champaign
Cloud Computing Concepts: Part 2
176 calificaciones
University of Illinois at Urbana-Champaign
176 calificaciones
Curso 2 de 6 en SpecializationCloud Computing
De la lección
Week 2: Concurrency and Replication Control
Lesson 1: Transactions are an important component of many cloud systems today. This module presents building blocks to ensure transactions work as intended, from Remote Procedure Calls (RPCs), to serial equivalence for transactions, to optimistic and pessimistic approaches to concurrency control, to deadlock avoidance/prevention. Lesson 2: This module covers how replication – maintaining copies of the same data at different locations – is used to provide many nines of availability in distributed systems, as well as different techniques for replication and for ensuring transactions commit correctly in spite of replication.
Conoce a los instructores
Indranil GuptaProfessor
Department of Computer Science
[MUSIC]
Continuing our discussion of replication control,
in this lecture we will see an important technique for
committing transactions in replication, in replicated systems.
This is known as two-phase commit.
In general, committing transactions is an important problem in replicated systems,
and in general in systems where you have objects across distributed servers.
So here's an example of a transaction that touches objects at multiple servers.
Here in this case, we ignore replication for now.
But let's just say each object has only one replica.
So there is 26 objects here.
There is 13 servers, each storing toward the objects.
So server 1 Stores objects A and B.
Servers two stores objects C and D.
And so on and so forth.
Server 13 stores objects Y and Z.
The transaction T here writes to all these objects.
When it reaches its commit point,
we need to make sure that all of these servers are commit.
So, when T tries to commit,
we need to make sure that all the servers commit their updates.
In this case, T will actually commit, the transaction T will commit.
Or that none of these servers commit to their updates to their objects.
In this case, T will abort.
Okay. So this is an example of an all or
nothing property here, which we want to guarantee for the transaction.
But we also want to guarantee on the server side in a group of servers
with distributed objects.
This problem, obviously, should be familiar to you.
This is, in fact, nothing but the consensus problem.
In fact, in the concept of transactions,
it's known as the atomic commit problem here.
So, one way to do the atomic commit is known as one-phase commit.
This is a first-cut naive approach where you have a coordinator server,
say, a special server elected using lead election protocol.
You are election protocol.
The coordinator server is a special server that issues on the commit and
simply tell the 13 servers involved here, hey,
why don't you commit your object updates from transaction T?
If something goes wrong on the transaction side, if the transaction T wants to abort,
then it tells that coordinator server to abort.
And it then, the coordinator server, then relays to the servers that
they should abort their updates from transaction T to their objects.
So what's a problem with this?
Well, the first issue is that the server, the server storing the objects,
the servers one through 13, have no say in whether the transaction commits or aborts.
These servers are, after all, machines.
So, they have the memory, and the memory might be corrupted.
For instance, the object that they're storing, or
the updates from that particular transaction may be corrupted.
And in this case, you'll end up with an inconsistent system because,
as far as the transaction T and the coordinator are concerned,
they have wanted to commit but the server just cannot commit.
It just cannot commit because it does not have all the necessary information there.
Also, the server might crash before receiving the message that asks it commit.
And in this case,
the updates it had from transaction T might have been stored in memory and lost.
And in this case, you, again, have a system where you have data being lost.
So, to our rescue comes a two-phase commit protocol, which to this day is one of
the most widely used protocols for committing transactions.
Here, again, you have a coordinator server.
Again, elected using a lead election protocol.
And you have the remaining servers.
As the name suggests, instead of using just one phase
where the coordinator tells the server, server group, you have two phases here.
So, the coordinator first sends a prepare message to the servers.
The prepare message asks the servers to prepare to commit.
When a server receives the prepare message, it saves the updates from that
particular transaction to disk so that is durable now.
And after this,
if the server fails, it can then recover starting from this log on disk.
Tthen, if everything is okay on the server side, it responds with a yes vote.
If something went wrong, for instance, the memory got corrupted or
it has lost its updates, it responds with a no vote.
The coordinator server waits for these votes.
If it receives any no vote, at least one no vote from any of the servers,
it will then send out an abort message to all the servers.
Okay? So, this ensures that,
unless all the servers have voted yes, you will, in fact,
end up aborting the entire transaction.
Also, the coordinator server might not receive all the votes within a time out.
If it does not receive all the votes within a time out, it also,
again, sends an abort message to everyone.
This is accounting for the case where a server might have failed,
server 13 might have failed, and might never be able to send a vote back.
And in this case,
the coordinator must abort because it doesn't know what happened,
what went wrong with server 13, and whether server 13 will ever come back up.
Okay, so once again, in summary, if the coordinator
server receives at least one no vote or it does not receive all the,
in this case 13 votes, within the time out, it will send an abort message.
However, if it receives all of the 13 votes within the time out, it then sends
out a commit message knowing that everything has gone right on the server
sides, on the server side, and they have saved their updates to the log on disk.
Once they have voted yes and
no, they can't immediately go ahead and commit, right?
So, if they voted yes, they can't really commit before receiving the next message.
And so, they need to wait until the second phase commit message is received.
When the commit message is received,
they can then commit updates from disk to the store, to the actual database, and
they can respond back with an acknowledgement back to the coordinator.
When the coordinator receives all the acknowledgements, okay,
it can then say that the two-phase commit is done.
How do you handle failures and how do you make sure the correctness is guaranteed?
So, first, if a server, one of the 13 servers in our case,
voted a yes, it cannot unilaterally go ahead and
commit the updates from the disk into the store or into the database.
It needs to wait for the commit message.
Why is this?
Well, one of the other servers in the group might have voted a no or
might have failed.
In this case, the server might have to abort in the future.
However, if a server votes no, meaning something has gone wrong on its site,
if it votes no in the very first phase, then it can abort right away.
Why is this?
Well, this is because the coordinator would have received at least one no vote.
And so, it would have decided an abort in the future.
So no matter when that abort is decided,
it will never be a commit because this server has voted a no.
Okay, so, if a server votes no, it can go ahead and abort right away.
But if a server votes yes, it needs to wait until the commit message is received.
How do you deal with crashes?
So, to deal with server crashes,
each server saves the tentative updates into permanent storage,
meaning disk, right before replying yes or no in the first phase.
It can't save it after replying yes or no because it might fail
right after replying yes or no, but before saving it on disk.
It needs to save it on disk before replying yes or no.
If it fails before this, before it saves it on disk,
anyway the coordinator will time out waiting for the server's response and
will abort the transaction.
If it times out, if it crashes after saving it on disk but
before replying yes or no, again, that's going to be an abort decision.
If it crashes after saving it on disk and sending a yes vote, then it might be
a commit decision, but at least it has all the necessary information saved on disc so
that it can recover or achieve it after it's recovery, after it has rebooted.
The coordinator might also crash.
And so, to recover from these crashes the coordinator
also logs all of it's decisions and the received and
sent messages including counts of how many okay's it received on disk so
that if it fails it can then recover and start where it left off.
However things might go wrong and the coordinator might never come back up,
so a new election might be required to elect a new coordinator.
In this case, the coordinator might need to either start where the old
two-phase commit left off by creating the servers,
or just restart the two-phase commit altogether again.
What about message losses?
The messages that are sent out during the two-phase commit might also get lost.
The prepare message,
which is the first state message going out from the coordinator to the servers
asking them to prepare to commit might get lost.
Now, if the prepare message gets lost, then the server, one of the 13 servers,
might decide to abort unilaterally after a timeout waiting for
the first phase message.
In this case, because it will never vote yes to the coordinator,
the coordinator will always decide an abort for this particular transaction.
Alternatively, to deal with prepare message loss you can have the coordinator
send out a second prepare message.
Now the yes no vote in the first phase, again, might get lost as well.
If this gets lost, then the coordinator will end up aborting the transaction.
This is tough luck.
Even though everything might have gone right, everyone might have voted yes,
just because the message, one of these yes votes was drawn by the network,
you decide to abort the transaction.
Yes, this is pessimistic.
But at least it maintains correctness, it errs on the side of safety.
And in this case, the coordinator, unless it receives all the votes,
will decide to abort the transaction.
Finally, the commit or abort message.
The second-phase message might get lost in the network itself.
And in this case, you're okay because the coordinator has already made a decision,
it has logged the decision on disk.
So, the server can poll the coordinator unless,
if it has not received a decision in the second phase for what its decision is.
Okay? So, it can poll it repeatedly until
the coordinator comes back up and sends it either a comment or an abort message.
So, a two-phase commit is used at the commit point in distributed servers.
Another alternative to two-phased comment is, of course,
a consensus solution such as Paxos.
So Paxos is also being widely used nowadays, increasingly used nowadays,
to decide whether to commit a transaction or not.
Okay, but if you use Paxos, you need to make sure that if any server votes a no,
then everyone will abort.
That's the only catch in using Paxos.
Paxos can also be used for
ordering updates that are being received by replica groups.
So, when a replica group receives multiple updates from multiple clients, you need
to make sure that all the replicas receive the updates in the same order.
This is, again, because in replicated state machines you want to make sure
that all the replicas have the same series of states at the transition through, and
the same series of outputs.
So, in order to do this, essentially, the way you use Paxos is that a server,
any one of the servers in the replica group, proposes a message for
the next sequence number.
So, one of the updates it has, it proposes it for the next sequence number.
And then, the group either reaches consensus on this or not.
If it doesn't, then another message is considered for
the next sequence number, and you continue so on and so forth.
Okay? So, this is the way in which
Paxos is used to totally order updates within the group itself so
that all the servers apply these updates in the same order as each other.
So that wraps up our discussion of concurrency control and
replication control.
And for replication control, I've discussed the case where you have multiple
servers, distributed servers in the cloud, which are either replicating objects for
fault-tolerance or load-balancing objects across multiple servers with or
without fault tolerance.
Replication can be implemented using a couple of flavors, active replication,
passive replication.
And these can be implemented using concepts that we've learned earlier in
the course, specifically, using flavors of multi-cast and virtual synchrony.
And, finally, we have seen how to commit transactions when you have distributed
servers in the Cloud.
The two phase commit protocol is one of the most widely used protocols today.
[MUSIC]
Explora nuestro catálogo
Inscríbete de manera gratuita y obtén recomendaciones personalizadas, actualizaciones y ofertas.
Coursera brinda acceso universal a la mejor educación del mundo, al asociarse con las mejores universidades y organizaciones, para ofrecer cursos en línea.
© 2018 Coursera Inc. Todos los derechos reservados.
