Related
I want to use consul for a 2-node cluster. Drawback is there's no failure tolerance for two nodes :
https://www.consul.io/docs/internals/consensus.html
Is there a way in Consul to make a consistent leader election with only two nodes? Can Consul Raft Consensus algorithm be changed?
Thanks a lot.
It sounds like you're limited to 2 machines of this type, because they are expensive. Consider acquiring three or five cheaper machines to run your orchestration layer.
To answer protocol question, no, there is no way to run a two-node cluster with failure tolerance in Raft. To be clear, you can safely run a two-node cluster just fine - it will be available and make progress like any other cluster. It's just when one machine goes down, because your fault tolerance is zero you will lose availability and no longer make no progress. But safety is never compromised - your data is still persisted consistently on these machines.
Even outside Raft, there is no way to run a two-node cluster and guarantee progress upon a single failure. This is a fundamental limit. In general, if you want to support f failures (meaning remain safe and available), you need 2f + 1 nodes.
There are non-Raft ways to improve the situation. For example, Flexible Paxos shows that we can require both nodes for leader election (as it already is in Raft), but only require a single node for replication. This would allow your cluster to continue working in some failure cases where Raft would have stopped. But the worst case is still the same: there are always failures that will cause any two-node cluster to become unavailable.
That said, I'm not aware of any practical flexible paxos implementations anyway.
Considering the expense of even trying to hack up a solution to this, your best bet is to either get a larger set of cheaper machines, or just run your two-node cluster and accept unavailability upon failure.
Talking about changing the protocol, there is impossibility proof by FLP which states that consensus cannot be reached if systems are less than 2f + 1 for f failures (fail-stop). Although, safety is provided but progress (liveness) cannot be ensured.
I think, the options suggested in earlier post are the best.
The choice of leader election on top of the Consul’s documentation itself requires 3 nodes. This relies on the health-checks mechanism, as well as the sessions. Sessions are essentially distributed locks automatically released by TTL or when the service crashes.
To build 2-node Consul cluster we have to use another approach, supposedly called Leader Lease. Since we already have Consul KV-storage with CAS support, we can simply write to it which machine is the leader before the expiration of such and such time. As long as the leader is alive and well, it can periodically extend it's time. If the leader dies, someone will replace it quickly. For this approach to work, it is enough to synchronize the time on the machines using ntpd and when the leader performs any action, verify that it has enough time left to complete this action.
A key is created in the KV-storage, containing something like “node X is the leader before time Y”, where Y is calculated as the current time + some time interval(T). As a leader, node X updates the record once every T/2 or T/3 units of time, thereby extending it's leadership role. If a node falls or cannot reach the KV-storage, after the interval(T) its place will be taken by the node, which will be the first to discover that the leadership role has been released.
CAS is needed to prevent a race condition if the two nodes simultaneously try to become a leader. CAS Specifies to use a Check-And-Set operation. This is very useful as a building block for more complex synchronization primitives. If the index is 0, Consul will only put the key if it does not already exist. If the index is non-zero, the key is only set if the index matches the ModifyIndex of that key.
Apache Kylin looks like a great tool that will fill in the needs of a lot data scientists. It's also a very complex system. We are developing an in-house solution with exactly the same goal in mind, multidimensional OLAP cube with low query latency.
Among the many issues, the one I'm concerned of the most right now is about fault tolerance.
With large volumes of incoming transactional data, the cube must be incrementally updated, and some of the cuboids are updated over long period of time such as those with time dimension value at the scale of year. Over such long period, some piece of the complex system is guaranteed to fail, and how does the system ensure all the raw transactional records are aggregated into the cuboids exactly once, no more no less? Even each of the pieces has its own fault tolerance mechanism, it doesn't mean they will play together automatically.
For simplicity, we can assume all the input data are saved in HDFS by another process, and can be "played back" in any way you want to recover from any interruption, voluntary or forced. What are Kylin's fault tolerance considerations, or is it not really an issue?
There are data faults and system faults.
Data fault tolerance: Kylin partitions cube into segments and allows rebuild an individual segment without impacting the whole cube. For example, assume a new daily segment is built on daily basis and get merged into weekly segment on weekend; weekly segments merge into monthly segment and so on. When there is data error (or whatever change) within a week, you need to rebuild only one day's segment. Data changes further back will require rebuild a weekly or monthly segment.
The segment strategy is fully customizable so you can balance the data error tolerance and query performance. More segments means more tolerable to data changes but also more scans to execute for each query. Kylin provides RESTful API, an external scheduling system can invoke the API to trigger segment build and merge.
A cube is still online and can serve queries when some of its segments is under rebuild.
System fault tolerance: Kylin relies on Hadoop and HBase for most system redundancy and fault tolerance. In addition to that, every build step in Kylin is idempotent. Meaning you can safely retry a failed step without any side effect. This ensures the final correctness, no matter how many fails and retries the build process has been through.
(I'm also Apache Kylin co-creator and committer. :-)
Notes: I'm Apache Kylin co-creator and committer.
The Fault Tolerance point is really good one which we actually be asked from some cases, when they have extreme large datasets. To calculate again from begin will require huge computing resources, network traffic and time.
But from product perspective, the question is: which one is more important between precision result and resources? For transaction data, I believe the exactly number is more important, but for behavior data, it should be fine, for example, the distinct count value is approximate result in Kylin now. It depends what's kind of case you will leverage Kylin to serve business needs.
Will put this idea into our backlog and will update to Kylin dev mailing list if we have more clear clue for this later.
Thanks.
Dynamo-like databases (e.g. Cassandra) can enforce consistency by means of quorum, i.e. a number of synchronously written replicas (W) and a number of replicas to read (R) should be chosen in such a way that W+R>N where N is a replication factor. On the other hand, PAXOS-based systems like Zookeeper are also used as a consistent fault-tolerant storage.
What is the difference between these two approaches? Does PAXOS provide guarantees that are not provided by W+R>N schema?
Yes, Paxos provides guarantees that are not provided by the Dynamo-like systems and their read-write quorums. The difference is how failures are handled and what happens during a write. After a successful write, both kind of systems behave similarly. The data will be saved and available for reading afterwards (until overwritten or deleted) and so on.
The difference appears during a write and after failures. Until you get a successful answer from W nodes when writing something to the eventually consistent systems, then the data may have been written to some nodes and not to others and there is no guarantee that the whole system agrees on the current value. If you try to read the data back at this point, some clients may get the new data back and some the old data back. In other words, the system is not immediately consistent. This is because writes aren't atomic across nodes in these systems. There are usually mechanisms to "heal" an inconsistency like this and "eventually" the system will become consistent again (i.e. reads will once again always return the same value, until something new is written). This is the reason why they are often called "eventually consistent". Inconsistencies can (and will) appear, but they will always be dealt with and reconciled eventually.
With Paxos, writes can be made atomic across nodes and inconsistencies between nodes are therefore possible to avoid. The Paxos algorithm makes it possible to guarantee that non-faulty nodes never disagree on the outcome of a write, at any point in time. Either the write succeeded everywhere or nowhere. There will never be any inconsistent reads at any point (if it's correctly implemented and if all the assumptions hold, of course). This comes at a cost, however. Mainly, the system may need to delay some requests and be unavailable when for example too many nodes (or the communication between them) aren't working. This is necessary to assure that no inconsistent replies are given.
To summarize: the main difference is that the Dynamo-like systems can return inconsistent results during writes or after failures for some time (but will eventually recover from it), whereas Paxos based systems can guarantee that there are never any such inconsistencies by sometimes being unavailable and delaying requests instead.
Paxos is non-trivial to implement, and expensive enough that many systems using it use hints as well, or use it only for leader election, or something. However, it does provide guaranteed consistency in the presence of failures - subject of course to the limits of its particular failure model.
The first quorum based systems I saw assumed some sort of leader or transaction infrastructure that would ensure enough consistency that you could trust that the quorum mechanism worked. This infrastructure might well be Paxos-based.
Looking at descriptions such as https://cloudant.com/blog/dynamo-and-couchdb-clusters/, it would appear that Dynamo is not based on an infrastructure that guarantees consistency for its quorum system - so is it being very clever or cutting corners? According to http://muratbuffalo.blogspot.co.uk/2010/11/dynamo-amazons-highly-available-key.html, "The Dynamo system emphasizes availability to the extent of sacrificing consistency. The abstract reads "Dynamo sacrifices consistency under certain failure scenarios". Actually, later it becomes clear that Dynamo sacrifices consistency even in the absence of failures: Dynamo may become inconsistent in the presence of multiple concurrent write requests since the replicas may diverge due to multiple coordinators." (end quote)
So, it would appear that in the case of quorums as implemented in Dynamo, Paxos provides stronger reliability guarantees.
Paxos and the W+R>N quorum try to solve slightly different problems. Paxos is usually described as a way to replicate a state machine, but in fact it is more of a distributed log: each item written to the log gets an index, and the different servers eventually hold the same log items + their index. (Replicated state machine can be achieved by writing to the log the inputs to the state machine and each server replays the state machine on the agreed inputs according to their index). You can read more about Paxos in a blog post I wrote here.
The W+R>N quorum solves the problem of sharing a single value among multiple servers. In the academia it is called "shared register". A shared register has two operations: read and write, where we expect the read to return the value of the previous write.
So, Paxos and the W+R>N quorum live in different domains, and have different properties (e.g., Paxos saves an ordered list of items). However, Paxos can be used to implement a shared register, and a W+R>N quorum can be used to implement a distributed log (although, very inefficiently).
Saying all the above, sometimes the W+R>N quorums aren't implemented in their "fully robust" way, as it will require more than one communication round. Thus, in systems that want low latency, it is possible that their implementation of W+R>N quorums provide weaker properties (e.g., conflicting values can co exist).
To sum up, theoretically, Paxos and the W+R>N can achieve the same goals. Practically, it would be very inefficient, and each one is better for something slightly different. Even more practically, W+R>N isn't always implemented fully, thus scarifying some consistency properties for speed.
Update: Paxos supports a very general failure model: messages can be dropped, nodes can crash and restart. The W+R>N quorum scheme has dfferent implementations, many of which assume less general failures. So, the difference between the two also depends on the assumption on the possible failures that are supported.
There is no difference. The definition of a quorum says that any two quorums' intersection is not empty. Simple majority quorum is an example NOT a definition. Take a look at Dr. Lamport's later paper "Vertical Paxos", where he gave some other possible configuration of quorums.
Multi-decree paxos protocol (AKA Multi-Paxos), in steady state it's just two phase commit. Ballot number changes are only needed when the leader fails.
Zookeeper's replication protocol (ZAB) , and RAFT are all based on Paxos. The differences are in fault-detection and transition after a leader fails.
As mentioned in other answers, in an R+W > N system, the writes are not atomic on all nodes which means that when a write is in progress (or during a write failure) some nodes will have newer values and some older ones. Take an example of a system where n=3, r=2, and w=2. For clarity let's assume the 3 nodes are named A, B, and C. Consider this scenario: a write is in progress; node A has been updated while B and C are still in process of receiving the updated value. Clients reading from A and B will see the newer value (resolved using version vectors or last write wins) and clients reading from B and C will see old values. This type of read is not considered linearizable. Such issues will not occur with proper linearizable systems such as Paxos or Raft.
Could someone give me a list of real use cases of Paxos. That is real problems that require consensus as part of a bigger problem.
Is the following a use case of Paxos?
Suppose there are two clients playing poker against each other on a poker server. The poker server is replicated. My understanding of Paxos is that it could be used to maintain consistency of the inmemory data structures that represent the current hand of poker. That is, ensure that all replicas have the exact same inmemory state of the hand.
But why is Paxos necessary? Suppose a new card needs to be dealt. Each replica running the same code will generate the same card if everything went correct. Why can't the clients just request the latest state from all the replicated servers and choose the card that appears the most. So if one server had an error the client will still get the correct state from just choosing the majority.
You assume all the servers are in sync with each other (i.e., have the same state), so when a server needs to select the next card, each of the servers will select the exact same card (assuming your code is deterministic).
However, your servers' state also depends on the the user's actions. For example, if a user decided to raise by 50$ - your server needs to store that info somewhere. Now, suppose that your server replied "ok" to the web-client (I'm assuming a web-based poker game), and then the server crashed. Your other servers might not have the information regarding the 50$ raise, and your system will be inconsistent (in the sense that the client thinks that the 50$ raise was made, while the surviving servers are oblivious of it).
Notice that majority won't help here, since the data is lost. Moreover, suppose that instead of the main server crashing, the main server plus another one got the 50$ raise data. In this case, using majority could even be worse: if you get a response from the two servers with the data, you'll think the 50$ raise was performed. But if one of them fails, then you won't have majority, and you'll think that the raise wasn't performed.
In general, Paxos can be used to replicate a state machine, in a fault tolerant manner. Where "state machine" can be thought of as an algorithm that has some initial state, and it advances the state deterministically according to messages received from the outside (i.e., the web-client).
More properly, Paxos should be considered as a distributed log, you can read more about it here: Understanding Paxos – Part 1
Update 2018:
Mysql High Availability uses paxos: https://mysqlhighavailability.com/the-king-is-dead-long-live-the-king-our-homegrown-paxos-based-consensus/
Real world example:
Cassandra uses Paxos to ensure that clients connected to different cluster nodes can safely perform write operations by adding "IF NOT EXISTS" to write operations. Cassandra has no master node so two conflicting operations can to be issued concurrently at multiple nodes. When using the if-not-exists syntax the paxos algorithm is used order operations between machines to ensure only one succeeds. This can then be used by clients to store authoritative data with an expiration lease. As long as a majority of Cassandra nodes are up it will work. So if you define the replication factor of your keyspace to be 3 then 1 node can fail, of 5 then 2 can fail, etc.
For normal writes Caassandra allows multiple conflicting writes to be accepted by different nodes which may be temporary unable to communicate. In that case doesn't use Paxos so can loose data when two Writes occur at the same time for the same key. There are special data structures built into Cassandra that won't loose data which are insert-only.
Poker and Paxos:
As other answers note poker is turn based and has rules. If you allow one master and many replicas then the master arbitrates the next action. Let's say a user first clicks the "check" button then changes their mind and clicks "fold". Those are conflicting commands only the first should be accepted. The browser should not let them press the second button it will disable it when they pressed the first button. Since money is involved the master server should also enforce the rules and only allow one action per player per turn. The problem comes when the master crashes during the game. Which replica can become master and how do you enforce that only one replica becomes master?
One way to handle choosing a new master is to use an external strong consistently service. We can use Cassandra to create a lease for the master node. The replicas can timeout on the master and attempt to take the lease. As Cassandra is using Paxos it is fault tolerant; you can still read or update the lease even if Cassandra nodes crash.
In the above example the poker master and replicas are eventually consistent. The master can send heartbeats so the replicas know that they are still connected to the master. That is fast as messages flow in one direction. When the master crashes there may be race conditions in replicas trying to be the master. Using Paxos at that point gives you strong consistently on the outcome of which node is now the master. This requires additional messages between nodes to ensure a consensus outcome of a single master.
Real life use cases:
The Chubby lock service for loosely-coupled distributed systems
Apache ZooKeeper
Paxos is used for WAN-based replication of Subversion repositories and high availability of the Hadoop NameNode by the company I work for (WANdisco plc.)
In the case you describe, you're right, Paxos isn't really necessary: A single central authority can generate a permutation for the deck and distribute it to everyone at the beginning of the hand. In fact, for a poker game in general, where there's a strict turn order and a single active player as in poker, I can't see a sensible situation in which you might need to use Paxos, except perhaps to elect the central authority that shuffles decks.
A better example might be a game with simultaneous moves, such as Jeopardy. Paxos in this situation would allow all the servers to decide together what sequence a series of closely timed events (such as buzzer presses) occurred in, such that all the servers come to the same conclusion.
Assuming I have a cluster of n Erlang nodes, some of which may be on my LAN, while others may be connected using a WAN (that is, via the Internet), what are suitable mechanisms to cater for a) different bandwidth availability/behavior (for example, latency induced) and b) nodes with differing computational power (or even memory constraints for that matter)?
In other words, how do I prioritize local nodes that have lots of computational power, over those that have a high latency and may be less powerful, or how would I ideally prioritize high performance remote nodes with high transmission latencies to specifically do those processes with a relatively huge computations/transmission (that is, completed work per message ,per time unit) ratio?
I am mostly thinking in terms of basically benchmarking each node in a cluster by sending them a benchmark process to run during initialization, so that the latencies involved in messasing can be calculated, as well as the overall computation speed (that is, using a node-specific timer to determine how fast a node terminates with any task).
Probably, something like that would have to be done repeatedly, on the one hand in order to get representative data (that is, averaging data) and on the other hand it might possibly even be useful at runtime in order to be able to dynamically adjust to changing runtime conditions.
(In the same sense, one would probably want to prioritize locally running nodes over those running on other machines)
This would be meant to hopefully optimize internal job dispatch so that specific nodes handle specific jobs.
We've done something similar to this, on our internal LAN/WAN only (WAN being for instance San Francisco to London). The problem boiled down to a combination of these factors:
The overhead in simply making a remote call over a local (internal) call
The network latency to the node (as a function of the request/result payload)
The performance of the remote node
The compute power needed to execute the function
Whether batching of calls provides any performance improvement if there was a shared "static" data set.
For 1. we assumed no overhead (it was negligible compared to the others)
For 2. we actively measured it using probe messages to measure round trip time, and we collated information from actual calls made
For 3. we measured it on the node and had them broadcast that information (this changed depending on the load current active on the node)
For 4 and 5. we worked it out empirically for the given batch
Then the caller solved to get the minimum solution for a batch of calls (in our case pricing a whole bunch of derivatives) and fired them off to the nodes in batches.
We got much better utilization of our calculation "grid" using this technique but it was quite a bit of effort. We had the added advantage that the grid was only used by this environment so we had a lot more control. Adding in an internet mix (variable latency) and other users of the grid (variable performance) would only increase the complexity with possible diminishing returns...
The problem you are talking about has been tackled in many different ways in the context of Grid computing (e.g, see Condor). To discuss this more thoroughly, I think some additional information is required (homogeneity of the problems to be solved, degree of control over the nodes [i.e. is there unexpected external load etc.?]).
Implementing an adaptive job dispatcher will usually require to also adjust the frequency with which you probe the available resources (otherwise the overhead due to probing could exceed the performance gains).
Ideally, you might be able to use benchmark tests to come up with an empirical (statistical) model that allows you to predict the computational hardness of a given problem (requires good domain knowledge and problem features that have a high impact on execution speed and are simple to extract), and another one to predict communication overhead. Using both in combination should make it possible to implement a simple dispatcher that bases its decisions on the predictive models and improves them by taking into account actual execution times as feedback/reward (e.g., via reinforcement learning).