I am working on a project that involves many clients connecting to a server(servers if need be) that contains a bunch of graph info (node attributes and edges). They will have the option to introduce a new node or edge anytime they want and then request some information from the graph as a whole (shortest distance between two nodes, graph coloring, etc).
This is obviously quite easy to develop the naive algorithm for, but then I am trying to learn to scale this so that it can handle many users updating the graph at the same time, many users requesting information from the graph, and the possibility of handling a very large (500k +) nodes and possibly a very large number of edges as well.
The challenges I can foresee:
with a constantly updating graph, I need to process the whole graph every time someone requests information...which will increase computation time and latency quite a bit
with a very large graph, the computation time and latency will obviously be a lot higher (I read that this was remedied by some companies by batch processing a ton of results and storing them with an index for later use...but then since my graph is being constantly updated and users want the most up to date info, this is not a viable solution)
a large number of users requesting information which will be quite a load on the servers since it has to process the graph that many times
How do I start facing these challenges? I looked at hadoop and spark, but they seem have high latency solutions (with batch processing) or solutions that address problems where the graph is not constantly changing.
I had the idea of maybe processing different parts of the graph and indexing them, then keeping track of where the graph is updated and re-process that section of the graph (a kind of distributed dynamic programming approach), but im not sure how feasible that is.
Thanks!
How do I start facing these challenges?
I'm going to answer this question, because it's the important one. You've enumerated a number of valid concerns, all of which you'll need to deal with and none of which I'll address directly.
In order to start, you need to finish defining your semantics. You might think you're done, but you're not. When you say "users want the most up to date info", does "up to date" mean
"everything in the past", which leads to total serialization of each transaction to the graph, so that answers reflect every possible piece of information?
Or "everything transacted more than X seconds ago", which leads to partial serialization, which multiple database states in the present that are progressively serialized into the past?
If 1. is required, you may well have unavoidable hot spots in your code, depending on the application. You have immediate information for when to roll back a transaction because it of inconsistency.
If 2. is acceptable, you have the possibility for much better performance. There are tradeoffs, though. You'll have situations where you have to roll back a transaction after initial acceptance.
Once you've answered this question, you've started facing your challenges and, I assume, will have further questions.
I don't know much about graphs, but I do understand a bit of networking.
One rule I try to keep in mind is... don't do work on the server side if you can get the client to do it.
All your server needs to do is maintain the raw data, serve raw data to clients, and notify connected clients when data changes.
The clients can have their own copy of raw data and then generate calculations/visualizations based on what they know and the updates they receive.
Clients only need to know if there are new records or if old records have changed.
If, for some reason, you ABSOLUTELY have to process data server side and send it to the client (for example, client is 3rd party software, not something you have control over and it expects processed data, not raw data), THEN, you do have a bit of an issue, so get a bad ass server... or 3 or 30. In this case, I would have to know exactly what the data is and how it's being processed in order to make any kind of suggestions on scaled configuration.
Related
The provided CSV dataset categories look like the following:
DATE | Hardware Identifier | What Failed | Description of Failure | Action Taken
The complete data can be easily downloaded from the Dropbox service using this link: data.csv
The data is very systematic, the input is very consistent and nicely structured. This data comes from a Computer Failure Data Repository. Additional details can be found on this link at USENIX: PNNL
About the data:
There are somewhat little over 2800 entries of single failure events that were collected over 4 years. Each event is described by the exact date and time when the event took place, what Node in the system failed, what hardware component of that node failed.
About the system:
Consists of 980 nodes processing some heavy calculation for the Molecular Science Computing Facility. Each node is designated by its own, unique ID.
My question:
Is it possible to perform any meaningful Machine Learning technique on such dataset, that would, in the end, be capable of predicting future failures in the system?
For example, would it be possible to train the ML algorithm on the provided dataset in order to predict either:
What node might fail soon (based on Hardware Identifier field)
What (node-piece of hardware) combination might fail soon (based on Hardware Identifier and either What failed or Description of Failure field)
What kind of failure might occur next anywhere in the system (based on What Failed field)
To me, this sounds like a huge classification problem. For example, in the case of (node-piece of hardware that failed), there are several thousands of different possibilities (classes). Having in mind that there are only little over 2800 single failure events described in the table, I don't feel like this would work.
Also, I am confused about how I should feed the data into the algorithm. Should the only input to the algorithm be the DATE field (converted to numeric linear growing time)? That doesn't seem right. Is it possible to feed the algorithm somehow with the time variable combined with some history of recent failure events? Should I restructure data to feed the algorithm with time variable + failure history (that might be limited, for example, to the last 30 days, or to feed the whole failure history of the system)?
May I hear your opinion? Is it possible to train an algorithm from this dataset that could predict any of the above-mentioned failure events (like, i.e. What node will fail next) given some input of the system (I can only think of time as an input for now, but that sounds wrong).
Since I am just starting to get involved with the ML algorithms, my thinking on the topic is probably very narrow and limited, so please feel free to suggest if you feel I should take a completely different approach on this.
Before we go on, remember that these failures are generally considered fairly random, so any results you get will likely be fairly unreliable.
The main problem to consider, is that you have very little data compared to the amount of nodes, slightly less than 3 on average, which means that you have to use some incredibly simple models, that would not give you much advantage over a random guess, for you to even have any certainty in your variables (separate mean time between failure would not have a determinable error, if it is even calculateable). For this I would probably treat each node as a separate test point, and then train a tree based algorithm to try to predict when the last failure in the nodes sequence of failures is, but that also mean that it would only be applicable to a subset of the database. This might be able to vaguely predict whether the node will fall in the near future and what type it would most likely be, but it like be fairly close to the estimate of mean time to failure and most common failure for all nodes.
If you want some meaningful results, you will need to have some attributes of the nodes that you can do the machine learning on, such as hardware components and when they were installed, and then have that as input in the classification. Since the problem will likely behave fairly randomly, you would get more information from trying to solve the regression problem instead of the classification problem, since you can still get good precision on a probabilistic model, even though the classification itself would be highly uncertain.
I'm not sure exactly where (or even how exactly to ask) this question, so I'm hoping someone here can point me in the right direction.
I have a service that I'm building. That service has different objects in memory - each with it's own state. Whenever an object is created it loads the state from the database and hold it. When changes are made to the object they are also persistent to the database.
I would like to scale this service. I have looked at solutions such as akka.net (actor model) and they have a clustering solution. From what I've read, it synchronizes the state with something they call "gossip" where each node sends the state to the other node. I'm not sure that it really possible to convert my working application to akka.net at this point.
I'm wondering exactly how clusters keep state synced between different nodes (I get the gossip concept), what happens if I have machine A that receives a message and at the same time, machine B also receives a message - both change the same state of an object - that will make problems with data integrity between states. My only thought about this is to lock a shared resource, but that defeats the purpose of the cluster.
Keeping state in the database is also not an option since the database becomes a bottleneck and a single point of failure.
I can't seem to find any relevant reading materials online - but I'm also lacking the technical phrases I need to focus on.
In case it's relevant, I'm using .NET Core and c# for development.
Can anyone explain the concept of clustering, how it works and make sure nodes are at sync? or can point to the right direction?
You have a big problem. I think that the way you are thinking about the problem is a bigger problem. Let's go through some basics.
Clustering is used to solve big problems, much like the "eat an elephant" problem. You could to solve this problem design a unique bigger predator with a huge mouth. But history and paleontology has shown us that big predators are not easily sustained (they are expensive on the environment).
So to solve your problem, you could take a bigger stronger server.
Or, you could use clustering.
Clustering solves the "eat the elephant" problem in a very different way. Instead of sending a unique huge predator with a huge mouth to eat the elephant, it will use a concept of distributed and shared processing to eat it one bite at a time. When done properly, ants could eat the elephant. If there are enough of them and the circumstances are correct.
But notice in my example, ants are very small... A single ant will never carry the entire elephant. You could carry the entire elephant if all the ants worked together but then you run into concurrency and locking problems (you must coordinate the ants).
Ants have shown us a much better way to deal with this. They will take a piece of the elephant and deal with the problem in smaller chunks.
In your system you ask how you can sync data across nodes... My question would be why? If you are syncing data then you are mirroring and your problem becomes even bigger (you are cloning the elephant but can only eat the original).
The solution to your problem is to rethink the solution and see if you can break down the problem into smaller pieces.
In Akka and in the Actor pattern the best way to deal with problems is to use smaller "processes" (a single ant). While the process on its own is almost useless, when used in a large scale they can become very powerful. When the architecture is properly done you will notice that taking a flamethrower to ants will not defeat them... More ants will come, they will continue to work on the problem.
Copying and syncing data is not your solution, clustering it is. You must take your data and break it down to a point where you can give it to a single ant. If you can do this then you can use Akka. If this approach seems ludicrous then Akka is not for you.
But consider this... You obviously have concerns over your database backend - you don't want to increase IO and introduce a single point of failure. I would have to agree with you. But you need to rethink things. You could have database mirroring to remove the single point of failure but you are correct that this won't remove the bottleneck. So let's say that mirror removes the single point of failure... Now let's attack the bottleneck portion.
If you can split up your data into small enough chunks that ants can handle it then I would urge you to tell your ants to only report to the database when the data changes... You can read it once on initialization (you need a backend store, don't kid yourself, electricity can be quickly lost... it must be saved somewhere) but if you tell your ants to persist only changed data then you will remove all the queries from the equation which will drastically shift where the load is coming from. Once you only have updates, inserts and deletes to deal with... the whole landscape will be much simpler.
Clustering should be the solution for you, but only if you can take the concept of mirror away from your mind.
Cluster nodes can and will crash... But they can be respawned elsewhere to other nodes, so that you always have a quick system. Only when you deal with a crash or loss of a node/worker process/ant will you have to reload data...
Good luck... you have outlined a formidable problem that I have seen people with software engineering degrees fail at solving.
I've been reading Introduction to Algorithms and started to get a few ideas and questions popping up in my head. The one that's baffled me most is how you would approach designing an algorithm to schedule items/messages in a queue that is distributed.
My thoughts have lead me to browsing Wikipedia on topics such as Sorting,Message queues,Sheduling, Distributed hashtables, to name a few.
The scenario:
Say you wanted to have a system that queued messages (strings or some serialized object for example). A key feature of this system is to avoid any single point of failure. The system had to be distributed across multiple nodes within some cluster and had to consistently (or as best as possible) even the work load of each node within the cluster to avoid hotspots.
You want to avoid the use of a master/slave design for replication and scaling (no single point of failure). The system totally avoids writing to disc and maintains in memory data structures.
Since this is meant to be a queue of some sort the system should be able to use varying scheduling algorithms (FIFO,Earliest deadline,round robin etc...) to determine which message should be returned on the next request regardless of which node in the cluster the request is made to.
My initial thoughts
I can imagine how this would work on a single machine but when I start thinking about how you'd distribute something like this questions like:
How would I hash each message?
How would I know which node a message was sent to?
How would I schedule each item so that I can determine which message and from which node should be returned next?
I started reading about distributed hash tables and how projects like Apache Cassandra use some sort of consistent hashing to distribute data but then I thought, since the query won't supply a key I need to know where the next item is and just supply it...
This lead into reading about peer to peer protocols and how they approach the synchronization problem across nodes.
So my question is, how would you approach a problem like the one described above, or is this too far fetched and is simply a stupid idea...?
Just an overview, pointers,different approaches, pitfalls and benefits of each. The technologies/concepts/design/theory that may be appropriate. Basically anything that could be of use in understanding how something like this may work.
And if you're wondering, no I'm not intending to implement anything like this, its just popped into my head while reading (It happens, I get distracted by wild ideas when I read a good book).
UPDATE
Another interesting point that would become an issue is distributed deletes.I know systems like Cassandra have tackled this by implementing HintedHandoff,Read Repair and AntiEntropy and it seems to work work well but are there any other (viable and efficient) means of tackling this?
Overview, as you wanted
There are some popular techniques for distributed algorithms, e.g. using clocks, waves or general purpose routing algorithms.
You can find these in the great distributed algorithm books Introduction to distributed algorithms by Tel and Distributed Algorithms by Lynch.
Reductions
are particularly useful since general distributed algorithms can become quite complex. You might be able to use a reduction to a simpler, more specific case.
If, for instance, you want to avoid having a single point of failure, but a symmetric distributed algorithm is too complex, you can use the standard distributed algorithm of (leader) election and afterwards use a simpler asymmetric algorithm, i.e. one which can make use of a master.
Similarly, you can use synchronizers to transform a synchronous network model to an asynchronous one.
You can use snapshots to be able to analyze offline instead of having to deal with varying online process states.
A few years back, researchers announced that they had completed a brute-force comprehensive solution to checkers.
I have been interested in another similar game that should have fewer states, but is still quite impractical to run a complete solver on in any reasonable time frame. I would still like to make an attempt, as even a partial solution could give valuable information.
Conceptually I would like to have a database of game states that has every known position, as well as its succeeding positions. One or more clients can grab unexplored states from the database, calculate possible moves, and insert the new states into the database. Once an endgame state is found, all states leading up to it can be updated with the minimax information to build a decision trees. If intelligent decisions are made to pick probable branches to explore, I can build information for the most important branches, and then gradually build up to completion over time.
Ignoring the merits of this idea, or the feasability of it, what is the best way to implement such a database? I made a quick prototype in sql server that stored a string representation of each state. It worked, but my solver client ran very very slow, as it puled out one state at a time and calculated all moves. I feel like I need to do larger chunks in memory, but the search space is definitely too large to store it all in memory at once.
Is there a database system better suited to this kind of job? I will be doing many many inserts, a lot of reads (to check if states (or equivalent states) already exist), and very few updates.
Also, how can I parallelize it so that many clients can work on solving different branches without duplicating too much work. I'm thinking something along the lines of a program that checks out an assignment, generates a few million states, and submits it back to be integrated into the main database. I'm just not sure if something like that will work well, or if there is prior work on methods to do that kind of thing as well.
In order to solve a game, what you really need to know per a state in your database is what is its game-theoretic value, i.e. if it's win for the player whose turn it is to move, or loss, or forced draw. You need two bits to encode this information per a state.
You then find as compact encoding as possible for that set of game states for which you want to build your end-game database; let's say your encoding takes 20 bits. It's then enough to have an array of 221 bits on your hard disk, i.e. 213 bytes. When you analyze an end-game position, you first check if the corresponding value is already set in the database; if not, calculate all its successors, calculate their game-theoretic values recursively, and then calculate using min/max the game-theoretic value of the original node and store in database. (Note: if you store win/loss/draw data in two bits, you have one bit pattern left to denote 'not known'; e.g. 00=not known, 11 = draw, 10 = player to move wins, 01 = player to move loses).
For example, consider tic-tac-toe. There are nine squares; every one can be empty, "X" or "O". This naive analysis gives you 39 = 214.26 = 15 bits per state, so you would have an array of 216 bits.
You undoubtedly want a task queue service of some sort, such as RabbitMQ - probably in conjunction with a database which can store the data once you've calculated it. Alternately, you could use a hosted service like Amazon's SQS. The client would consume an item from the queue, generate the successors, and enqueue those, as well as adding the outcome of the item it just consumed to the queue. If the state is an end-state, it can propagate scoring information up to parent elements by consulting the database.
Two caveats to bear in mind:
The number of items in the queue will likely grow exponentially as you explore the tree, with each work item causing several more to be enqueued. Be prepared for a very long queue.
Depending on your game, it may be possible for there to be multiple paths to the same game state. You'll need to check for and eliminate duplicates, and your database will need to be structured so that it's a graph (possibly with cycles!), not a tree.
The first thing that popped into my mind is the Linda-style of a shared 'whiteboard', where different processes can consume 'problems' off the whiteboard, add new problems to the whiteboard, and add 'solutions' to the whiteboard.
Perhaps the Cassandra project is the more modern version of Linda.
There have been many attempts to parallelize problems across distributed computer systems; Folding#Home provides a framework that executes binary blob 'cores' to solve protein folding problems. Distributed.net might have started the modern incarnation of distributed problem solving, and might have clients that you can start from.
I've been developing an in-house DSP application (Java w/ hooks for Groovy/Jython/JRuby, plugins via OSGi, plenty of JNI to go around) in data flow/diagram style, similar to pure data and simulink. My current design is a push model. The user interacts with some source component causing it to push data onto the next component and so on until a end block (typically a display or file writer). There are some unique challenges with this design specifically when a component starves for input. There is no easy way to request more input. I have mitiated some of this with feedback control flow, ex an FFT block can broadcast that it needs more data to source block of it's chain. I've contemplated adding support for components to be either push/pull/both.
I'm looking for responses regarding the merits of push vs pull vs both/hybrid. Have you done this before? What are some of the "gotchas"? How did you handle them? Is there a better solution for this problem?
Some experience with a "mostly-pull" approach in a large-scale product:
Model: Nodes build a 1:N tree, i.e. each component (except the root) has 1 parent and 1..N children. Data flows almost exclusively from parent to children. Change notifications can originate from any node in the tree.
Implementation: All leafs are notified with the sending node's id and a "generation" counter. Leafs know which node path they depend on, so they know if they need to update. (Any other child node update algorithm would do, too, and might have been better in hindsight).
Leafs query their parent for current data, query bubbles up recursively. The generation counter is included, so the bubble-up stops at the originating node.
Advantages:
parent nodes don't need much/any information about their children. Data can be consumed by anyone - this allowed a generic approach to implementing some (initially not expected) non-UI functionality on top of the data intended for display
Child nodes can aggregate and delay updates (avoiding repaints sure beats fast painting)
inactive leafs do cause no data traffic at all
Disadvantages:
Incremental updates are expensive, as full data is published.
The implementation actually allows for different data packets to be requested (and
the generation counter could prevent unecessary data traffic), but the data packets initially designed are very large. Slicing them was an afterthought, but works ok.
You need a real good generation mechanism. The one initially implemented collided with initial updates (that need special handling - see "incremental updates") and aggregation of updates
the need for data travelling up the tree was greatly underestimated.
publish is cheap only when the node offers read-only access to current data. This might require additional update synchronization, though
sometimes you want intermediate nodes to update, even when all leafs are inactive
some leafs ended up implementing polling, some base nodes ended up relying on that. ugly.
Generally:
Data-Pull "feels" more native to me when data and processing layer should know nothing about the UI. However, it requires a complex change notificatin mechanism to avoid "Updating the universe".
Data-Push simplifies incremental updates, but only if the sender intimately knows the receiver.
I have no experience of similar scale using other models, so I can't really make a recommendation. Looking back, I see that I've mostly used pull, which was less of a hassle. It would be interesting to see other peoples experiences.
I work on a pure-pull image processing library. It's more geared to batch-style operations where we don't have to deal with dynamic inputs and for that it seems to work very well. Pull works especially well for large data sets and for threading: we scale linearly to at least 32 CPUs (depending on the graph being evaluated, of course, heh).
We have a GUI that allows leaves to be dynamic data sources (for example, a video camera delivering frames) and they are handled by throwing away and rebuilding the relevant parts of the graph on a change. This is cheap in our case, so the overhead isn't that high.