Closed. This question does not meet Stack Overflow guidelines. It is not currently accepting answers.
We don’t allow questions seeking recommendations for books, tools, software libraries, and more. You can edit the question so it can be answered with facts and citations.
Closed 2 years ago.
Improve this question
I've quite understood what the Raft is and implemented it in MIT6.824 distributed system. I also know what's the basic Paxos, I've not implemented this yet, so I can't grab all details of it. For Multi-Paoxs, I'm even more confused, i.e., WHY it can eliminate lots of Prepare RPC? I know the answer should be Multi-Paxos can have a fixed leader along with noMoreAccepted response from other peers to determine if reduce Prepare RPC. But, I can't get it in detailed level, why and how it works
I want to get more some recommendations, articles, sample code or anything that can help for Multi-Paxos,
I've read Paxos made live and Paxos made simple, those two papers can give me a basic idea about what's Paxos and how it works
I've watched https://www.youtube.com/watch?v=YbZ3zDzDnrw&t=3035s&ab_channel=DiegoOngaro several times as well, it's a great talk, but it does not involve too many details
To answer your specific question:
WHY it can eliminate lots of Prepare RPC?
In the paper Paxos Made Simple page 10 it says:
A newly chosen leader executes phase 1 for infinitely many instances of the consensus algorithm—in the scenario above, for instances 135–137 and all instances greater than 139.
That is saying that if a leader broadcasts Prepare(135,n) which is a prepare for instance 135 using ballot number n then it is valid that this can be defined as applying to all instances >=135 that are not yet fixed. We can reason that it is safe for any node to be "spamming" out prepare messages for an infinite number of the unfixed positions in our log stream. This is because for each position each acceptor uses the rules of Paxos for that position. We can compress that infinite set of prepare messages down to a single one that applies to all higher unfixed positions. We then eliminate all but one prepare message for the term of a stable leader. So it is fantastic optimisation.
You asked about any example code. I wrote an implementation of multi-paxos using functional programming in Scala that aims to be true to the paper Paxos Made Simple over at https://github.com/trex-paxos/trex. The core state is PaxosData, the message protocol is at the bottom of PaxosProtcol and the algorithm is a set of message matching functions in PaxosAlgorithm. The algorithm takes the current immutable state and an immutable message as input and outputs the next immutable state for the node. Common behaviours are written as partial functions that have full unit tests. These partial functions are composed into complete functions used by leaders, followers and candidate leaders. There is a write up at this blog.
It adds additional messages to the basic set as optimisations speed up log replication. Those involve some implementation details that Lamport does not get into in his paper. An example is that negative acknowledgements are used to pass information between nodes to try to avoid interrupting a stable leader due to only one failed network link between a node and the leader. TRex tries to keep those features to a minimum to create a basic but complete solution.
An answer that you might find helpful about Multi-Paxos is this one that discusses why Multi-Paxos is called that https://stackoverflow.com/a/26619261/329496
There is also this one about how the original Part-Time Parliament paper uses a leader and also describes a stable leader running multi-Paxos https://stackoverflow.com/a/46012211/329496
Finally, you might enjoy my defence of Paxos post The Trial Of Paxos Algorithm.
Related
Closed. This question is opinion-based. It is not currently accepting answers.
Want to improve this question? Update the question so it can be answered with facts and citations by editing this post.
Closed 2 years ago.
Improve this question
I'm an engineer currently developing Linux kernel-mode drivers and user-mode drivers. When I came across the theory of 12 Factor App, there is a strong voice echoing around my brain "THIS IS THE FUTURE OF DEVELOPING!".
And I kept wondering how to apply this method to Linux KMD and UMD design and developing since this theory is much too web-app based (I'm a part-time open-source web developer).
Current Developing language: C
Current Testing automation: Custom Implemented Python testing framework (Progress based, NO unit test)
Please give me some suggestions on this. Thanks and appreciated in advance.
As with most development guidelines, there is a gap between the guideline and the enforcement.
For example, in your "12 factor app" methodology, one of the factors is:
Codebase - One codebase tracked in revision control, many deploys
Which sounds great, and would really simplify things. Until you get to the point of utility libraries. You see, when you find you are reusing code across multiple projects, you probably want:
Independent build and release chains for the multiple projects.
This could mean two codebases, but the above states one codebase (perhaps one per project, perhaps one per company. Let's assume one per company first, which is easy to see as non-ideal; because, you would have commits unrelated to a project in the project's commit history. Ok, one per project, more sensible; but, what if projects need to share code? Like the libraries that format their communications and control the send / receive protocols? Well, we could create a third "protocol library" so that we have revisioning around the protocol; but, that violates the "one codebase (per project)" because now you have two codebases comprising the single releasable item.
The decisions here are not simple. The other approach is to copy the protocol code into both projects and keep them in sync by some other means.
Dependencies - Explicitly declare and isolate dependencies
It's a great idea; and, one that makes development easier in many ways. Again, just to illustrate how a great idea can suffer without clear guidelines on how to implement the idea, what do you do when you are using a library that doesn't attempt to isolate the dependencies the library uses? Many of the more complex libraries themselves depend on other libraries, and generally they clearly declare their dependencies, as do the libraries used by the libraries; however, sometimes the base, core libraries used by multiple projects (logging, configuration, etc) wind up being used at different release versions. The isolation occurred on a per-library basis, but not on a per-project basis. You could fix it, provided you wanted (or could) fork and clone the libraries, restructuring them to properly isolate their dependencies for overall coordination of version numbers; but, generally you will lack the time to work on other people's projects.
In general, the advice under "12 factor app" methodology is good; but, it leaves you up to performing the work of translating the guidelines into development protocols. Enforcement then becomes a matter of interpertation, and the means of enforcement (as well as the interpertation) fall on you to implement.
And some of the guidelines look dangerously over-simlpified:
Concurrency - Scale out via the process model
While this is an easier way to go, it's not how any single high performance web server works. They all use threading, thread pools, and other more complex constructs to avoid process switching. These constructs (which are admittedly harder to use) were created specifically due to the limitations of a traditional process model. After all, it's not common to launch a process per web request, nor would you generally "tune a program for better performance" by starting a second copy on the same machine. Certainly, there are architectures where this could work; but, so far these architectures haven't outperformed their competition.
Between machines, I wholeheartedly agree. Process scalaing is the only way to go in a distrubuted environment; but, there's not much in this methodology that talks about distributed algorithms, or even distributed computing approaches; so, again it's another thing left up to the implementor.
Finally, their process commentary seems really out-of-place for writing a command line tool. The push to daemonize things works really well for microservices; however, you can't microservice away even the clients. Eventually you'll have to write something that isn't "managed by systemd", due to starting execution and ending execution without having an always-on service.
So, it's a good framework, which might not work for some things, even if it is excellent for many things; but, in my opinion, the tooling to enforce it would have to be built by the organization using it because the interpretations one organization might make could differ from another organization.
Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 3 years ago.
Improve this question
I am preparing for the system design interview, and since I have little experience with this topic, I bought the "Grokking the system design interview" course from educative.io, which was recommended by several websites.
However I read it, I think I did not manage to understand several things, so if someone could answer my questions, that would be helpful.
Since I have no experience with nosql, I find it difficult to chose the proper db system. Several times the course just do not give any reasoning why it chose one db over another one. For example in chapter "Designing Youtube or Netflix" the editors chose mysql for db role with no explanation. In the same chapter we have the following non-functional requirements:
"The system should be highly available. Consistency can take a hit (in
the interest of availability); if a user doesn’t see a video for a
while, it should be fine."
Following the above hint and taking into account the size of the system and applying the material in the "CAP theorem" chapter for me it seems or Cassandra and CouchDB would be a better choise. What do I miss here?
Same question goes for "Designing Facebook’s Newsfeed"
Is CAP theorem still applicable?
What I mean is: HBase is according to the chapter "CAP theorem" good at consistency and partition tolerance, but according to the HBase documentation, it also supports High Availibility since version 2.X. So it seems to me that it is a one fits all / universal solution for db storage which goes against CAP theorem, unless they sacrificed something for HA. What do I miss here?
The numbers are kind of inconsistent around the course about how much RAM/storage/bandwidth can a computer handle, I guess they are outdated. What are the current numbers for a, regular computers, b, modern servers?
Almost every chapter has a part called "Capacity Estimation and Constraints", but what is calculated here changes from chapter to chapter. Sometimes only storage is calculated, often bandwidth too, sometimes QPS is added, sometimes there are task specific metrics. How do I know what should I calculate for a specific task?
Thanks in advance!
Each database is different and fulfills different requirements. I recommend you read dynamo-paper, and familiarize yourself with the rest of the terminology used in it (two-phase locking, leader/follower, multi-leader, async/sync replication, quorums), and know what guarantees the different databases provide. Now to the questions:
MySQL can be configured to prioritize Availability at the cost of Consistency with its asynchronous replication model (the leader doesn't wait for acknowledgement from its followers before committing a write; if a leader crashes before the data gets propagated to the followers, the data is lost), so it can be one of the suitable solutions here.
From the documentation of HBase, HBase guarantees strong consistency, even at the cost of availability.
The promise of high availability is for reads, not for writes i.e. for reading stale data while the rest of the system recovers from failure and can accept additional writes.
because of this single homing of the reads to a single location, if the server becomes unavailable, the regions of the table that were hosted in the region server become unavailable for some time.
Since all writes still have to go through the primary region, the writes are not highly-available (meaning they might block for some time if the region becomes unavailable).
The numbers used are estimates by the candidate i.e. you decide what are the specs of a single hypothetical server, and how many servers you would need in order to scale and accommodate the storage/throughput requirement.
You don't know in advance (although you can make a guess based on the requirements e.g. if it's a data storage system, a streaming service etc., I still wouldn't recommend it). Instead, you should ask the interviewer what area they are interested in, and you make estimates for it. The interview, especially the system design part, is a discussion, don't follow a template to the letter. You recognize the different areas you can tackle about the system, and approach them based on the interviewer's interest.
Closed. This question does not meet Stack Overflow guidelines. It is not currently accepting answers.
We don’t allow questions seeking recommendations for books, tools, software libraries, and more. You can edit the question so it can be answered with facts and citations.
Closed 3 years ago.
Improve this question
We have a web service that needs a somewhat POSIX-compatible shared filesystem for the application servers (multiple redundant systems running in parallel behind redundant load balancers). We're currently running GlusterFS as the shared filesystem for the application servers but I'm not happy with the performance of the system. Compared to actual raw performance of the storage servers running GlusterFS, it starts to look more sensible to run DRBD and single NFS server with all the other GlusterFS servers (currently 3 servers) waiting in hot-standby role.
Our workload is highly read oriented and usually deals with small files and I'd be happy to use "eventually consistent" system as long as a client can request sync for a single file if needed (that is, client is prepared to wait until the file has been successfully stored in the backend storage). I'd even accept a system where such "sync" requires querying the state of the file via some other way than POSIX fdatasync(). File metadata such as modification times is not important, only filename and the contents.
I'm currently aware of possible candidates and the problems each one currently has:
GlusterFS: overall performance is pretty poor in practice, performance goes down while adding new servers/bricks.
Ceph: highly complex to configure/administrate, POSIX compatibility sacrifices performance a lot as far as I know.
MooseFS: partially obfuscated open source (huge dumps of internally written code published seldomly with intentionally lost patch history), documentation leaves lots to desire.
SeaweedFS: pretty simple design and supposedly high performance, future of this project is unclear because pretty much all code is written and maintained by Chris Lu - what happens if he no longer writes any code? Unclear if the "Filer" component supports no single point of failure.
I know that CAP theorem prevents ever having truly consistent and always available system. Is there any good system for distributed file system where writes must be durable, but read performance is really good and the system has no single point of failure?
I am Chris Lu working on SeaweedFS. There are plans to commercialize it. (By adding more advanced features.)
The filer does not have simple point of failure, you can have multiple filer instances. The filer store can be any key-value store. If you need no SPOF, you can use Cassandra, Redis cluster, CockroachDB, TiDB, or Etcd. Or you can add your own key-value store option, which is pretty easy.
Closed. This question does not meet Stack Overflow guidelines. It is not currently accepting answers.
Closed 4 years ago.
This question does not appear to be about a specific programming problem, a software algorithm, or software tools primarily used by programmers. If you believe the question would be on-topic on another Stack Exchange site, you can leave a comment to explain where the question may be able to be answered.
We don’t allow questions seeking recommendations for books, tools, software libraries, and more. You can edit the question so it can be answered with facts and citations.
Improve this question
I have an AWS RDS (AuroraDB) and I want to mask the data on the DB. Does Amazon provides any service for data masking?
I have seen RDS encryption but I am looking for data masking because the database contains sensitive data. So I want to know is there any service they provide for data masking or is there any other tool which can be used to mask the data and add it manually into the DB?
A list of tools which can be used for data masking is most appreciated if any for mine case. Because I need to mask those data for testing as the original DB contains sensitive information like PII(Personal Identifiable information). I also have to transfer these data to my co-workers, so I consider data masking an important factor.
Thanks.
This is a fantastic question and I think your pro-active approach to securing the most valuable asset of your business is something that a lot of people should heed, especially if you're sharing the data with your co-workers. Letting people see only what they need to see is an undeniably good way to reduce your attack surfaces. Standard cyber security methods are no longer enough imo, demonstrated by numerous attacks/people losing laptops/usbs with sensitive data on. We are just humans after all. With the GDPR coming in to force in May next year, any company with customers in the EU will have to demonstrate privacy by design and anonymisation techniques such as masking have been cited as way to show this.
NOTE: I have a vested interest in this answer because I am working on such a service you're talking about.
We've found that depending on your exact use case, size of data set and contents will depend on your masking method. If your data set has minimal fields and you know where the PII is, you can run standard queries to replace sensitive values. i.e. John -> XXXX. If you want to maintain some human readability there are libraries such as Python's Faker that generate random locale based PII you can replace your sensitive values with. (PHP Faker, Perl Faker and Ruby Faker also exist).
DISCLAIMER: Straight forward masking doesn't guarantee total privacy. Think someone identifying individuals from a masked Netflix data set by cross referencing with time stamped IMDB data or Guardian reporters identifying a Judges porn preferences from masked ISP data.
Masking does get tedious as your data set increases in fields/tables and you perhaps want to set up different levels of access for different co-workers. i.e. data science get lightly anonymised data, marketing get a access to heavily anonymised data. PII in free text fields is annoying and generally understanding what data is available in the world that attackers could use to cross reference is a big task.
The service i'm working on aims to alleviate all of these issues by automating the process with NLP techniques and a good understanding of anonymisation maths. We're bundling this up in to a web-service and we're keen to launch on the AWS marketplace. So I would love to hear more about your use-case and if you want early access we're in private beta at the moment so let me know.
If you are exporting or importing data using CSV or JSON files (i.e. to share with your co-workers) then you could use FileMasker. It can be run as an AWS Lamdbda function reading/writing CSV/JSON files on S3.
It's still in development but if you would like to try a beta now then contact me.
Disclaimer: I work for DataVeil, the developer of FileMasker.
Closed. This question needs details or clarity. It is not currently accepting answers.
Want to improve this question? Add details and clarify the problem by editing this post.
Closed 6 years ago.
Improve this question
Using the api to analyze a twitter stream I am getting very similar results for openness for pretty much everybody. How can I train a corpus to generate a different output
Unfortunately, you can't. Also, I am afraid twitter is not the best source for this kind of analysis since each tweet has just a little piece of text. Watson Personality Insights works better with large text samples, and most probably, twitter sentences are too short to provide enough information for this kind of analysis (even if you concatenate several tweets in the same text sample).
But, if you're getting meaningful results for the other dimensions, what I'd suggest you to do is to ignore the openness information and try to calculate it using another algorithm (your own?) or even checking if just removing this dimension does not provide good enough results for you.
There are some nice tips here -- https://www.ibm.com/smarterplanet/us/en/ibmwatson/developercloud/doc/personality-insights/science.shtml and some references to papers that can help you understand the algorithm internals.
You cannot train Watson Personality Insights at the current version. But there may be alternatives.
From your message it is not clear to me if you are receiving too similar results for individual tweets or entire twitter streams. In the first case, as Leo pointed out in a different answer, please note that you should aim to provide enough information for any analysis to be meaningful (this is 3,000+ words, not just a tweet). In the second case, I would be a bit surprised if your scores are still so similar with so much text (how many tweets per user?), but this may still happen depending on the domain.
If you are analyzing individual tweets you may also benefit from user Tone Analyzer (in Beta as of today). Its "social tone" is basically the same model as Personality Insights, and gives some raw scores even for small texts. (And by the way you get other measures such as emotions and writing style).
And in any case (small or large inputs), we encourage users to take a look at the raw scores in their own data corpus. For example, say you are analyzing a set of IT support calls (I am making this up), you will likely find some traits tend to be all the same because the jargon and writing style is similar in all of them. However, within your domain there may be small differences you may want to focus, ie. there is still a 90% percentile, a lowest 10% in each trait... So you might want to do some data analysis on Personality Insights raw_score (api reference) or just the score in Tone Analyzer (api reference) and draw your own conclusions.