Would it be possible to build a Ticketmaster style ticket reservation system by storing all information in a Cassandra cluster?
The system needs to be able to
1. Display the correct number of tickets available at one time
2. Temporarily reserve a ticket while the customer is making the purchase
3. No two users can ever buy the same ticket.
For consistency all reads and writes should be made at quorum. I'm not sure how to implement steps 2 or 3?
Yes, you can.
However, there will be some transactions where you want strict consistency. For example, consistency does not matter when the user is browsing the site and adding tickets to their shopping cart, but when they checkout and select a specific seat number on a specific day consistency matters a great deal (double bookings being a bad thing, especially for high interest events).
So, you could implement 99% of the functionality in an eventually consistent database and implement the checkout process in a consistent database. This is also nice because you can scale 99% of your system that likely gets >70% of the load horizontally and across multiple data centers. Just keep in mind that you will have to deal with the scenario of your site being up but your checkout process being down (ex., an error dialog at checkout asking them to wait/retry and giving them a promo code for their troubles).
The last detail is that you will need to update your eventually consistent database's "number of available tickets" after someone checks out. The good news is that this can be done lazily - queue up that job and do it whenever your system has some spare cycles. It certainly never has to happen in the critical path of the user's checkout process.
Related
I am new in learning distributed systems and I read about the CAP theorem, I am interested in an AP system such as Cassandra.
My question is in what cases can you actually sacrifice consistency? Effectively what I am saying is sacrificing consistency means serving inaccurate data. In what cases would then you actually use an AP datastore like Cassandra? I can't think of any case where I wouldn't want my reads to be consistent.
By AP system, I assume you will at least target to ensure eventual consistency.
Imagine you're developing a social network where users have friends and their own news feeds. It doesn't matter if a particular user's feed has occasional five minutes lag (his feed list has eventual consistency). Missing 2/3 very recent updates in the news feed is okay in this scenario as long as those feeds will eventually appear. And in fact, Facebook built it's news feed using Cassandra.
Imagine a distributed key-value store cache system where update is very rare. If there is almost no update operations, ensuring strong consistency is un-necessary, so you can focus on availability. Occasional cache miss (the key-value entry is not populated yet) and request to database due to eventual consistency should be okay.
My question is in what cases can you actually sacrifice consistency?
One case would be when building a recommendation engine data set and serving it with Cassandra. These data sets are essentially the aggregation of many, many users to determine purchasing/viewing patterns.
For example: If I add a Rey Star Wars action figure to my shopping cart, the underlying recommendation engine runs a query for similar resulting purchasing patterns based on others who have also purchased an action figure of Rey. The query returns the top 5 product results, and puts them at the bottom of the page.
Those 5 products returned are the result of analysis and aggregation of several thousand prior purchases. Let's assume that some of that data isn't consistent, causing a variance in the 5 products returned. Is that really a big deal?
tl;dr; The real question to ask; is whether or not getting a somewhat-accurate list of 5 product recommendations in less than 10ms, is better than getting a 100% accurate list of 5 product recommendations in 100ms?
Both result sets will help drive sales. But the one which is returned fast enough that it doesn't hinder the user experience is much more preferred.
'C' in CAP refers to linearizability which is a very strong form of consistancy that you don't need most of the time.
Linearizability is a recency guarantee which makes it appear that there is a single copy of data. As soon as you make a change in the data, all subsequent reads will return the changed data. Such a level of consistency is expensive and doesn't scale well. Yet in certain scenarios we need linearizability, viz.
Leader election
Allowing end users to create their unique user id
Distributed locking etc.
When you have these usecases, you'd use something like ZooKeeper, etcd etc. Cassandra also has Light Weight Transaction (LWT) which uses an extension of the classic Paxos algorithm to implement linearizability. This feature can be used to address those rare use cases where you must have linearizability and serializability, but it is expensive. And in vast majority of cases you are just fine with a little weaker consistency to get better scalability and performance. You trade a little bit of consistency with scalability and performance.
Some eCommerce websites send apology letter to customers for not being able to fulfill their orders. That is because the last copy of the product has been sold to more than one customers due to lack and linearizability. They prefer to deal with that over not being able to scale with the customer base and not being able to respond to their requests within stringent SLAs.
Cassandra is said to have a tuneable consistency. You may want to record user clicks or activities for analysis. You are okay if some data are lost, but you cannot compromise with the performance. You'd probably use a write consistency level of ANY with hints enabled (sloppy quorum).
If you want a little more consistency, you'd use a QUORUM consistency level to read and write along with hints and read repair. In vast majority of case all nodes are updated instantaneously. Even if one or two nodes go down, a majority of nodes will have the data and failed nodes would be repaired when they come back using hints, read repair, anti entropy repair.
Cassandra is particularly useful for cases where you'd not have many concurrent updates on same data. The reason is, unlike the dynamo architecture, it does not use vector clocks for conflict resolution between replicas. Instead it uses Last Write Wins (LWW) based on timestamp. If timestamps are same, it uses lexicographical order. Since the time on nodes cannot be accurate even in the presence of NTPD, there is a possibility of data loss, although Cassandra has taken some steps to avoid that - for e.g. client side timestamp instead of server side timestamp.
The CAP theorem says that given partition tolerence, you can either choose availability or consistency in a distributed database (no one would want to give up partition tolerence in any case). So if you want to have maximum availability, you'll have to give up on the consistency. This depends of course, on how critical the business is.
You answered something on SO but the answer doesn't show up when you visit the page? Can be tolerated. SO being down? Can't be. Critical financial systems would rather have strong consistency than availability. Every once-in-a-while, my bank's servers would go offline when I try to make a payment.
Normally, you choose availability and eventual consistency. The answer you wrote into SO would eventually show up.
Apart from the above mentioned cases where inconsistent data is tolerable, there are also scenarios where we can defer to the user to solve the inconsistency.
For example, if we found two different versions of someone's address in the database, we can prompt the user to identity the correct address.
What is the best way to achieve DB consistency in microservice-based systems?
At the GOTO in Berlin, Martin Fowler was talking about microservices and one "rule" he mentioned was to keep "per-service" databases, which means that services cannot directly connect to a DB "owned" by another service.
This is super-nice and elegant but in practice it becomes a bit tricky. Suppose that you have a few services:
a frontend
an order-management service
a loyalty-program service
Now, a customer make a purchase on your frontend, which will call the order management service, which will save everything in the DB -- no problem. At this point, there will also be a call to the loyalty-program service so that it credits / debits points from your account.
Now, when everything is on the same DB / DB server it all becomes easy since you can run everything in one transaction: if the loyalty program service fails to write to the DB we can roll the whole thing back.
When we do DB operations throughout multiple services this isn't possible, as we don't rely on one connection / take advantage of running a single transaction.
What are the best patterns to keep things consistent and live a happy life?
I'm quite eager to hear your suggestions!..and thanks in advance!
This is super-nice and elegant but in practice it becomes a bit tricky
What it means "in practice" is that you need to design your microservices in such a way that the necessary business consistency is fulfilled when following the rule:
that services cannot directly connect to a DB "owned" by another service.
In other words - don't make any assumptions about their responsibilities and change the boundaries as needed until you can find a way to make that work.
Now, to your question:
What are the best patterns to keep things consistent and live a happy life?
For things that don't require immediate consistency, and updating loyalty points seems to fall in that category, you could use a reliable pub/sub pattern to dispatch events from one microservice to be processed by others. The reliable bit is that you'd want good retries, rollback, and idempotence (or transactionality) for the event processing stuff.
If you're running on .NET some examples of infrastructure that support this kind of reliability include NServiceBus and MassTransit. Full disclosure - I'm the founder of NServiceBus.
Update: Following comments regarding concerns about the loyalty points: "if balance updates are processed with delay, a customer may actually be able to order more items than they have points for".
Many people struggle with these kinds of requirements for strong consistency. The thing is that these kinds of scenarios can usually be dealt with by introducing additional rules, like if a user ends up with negative loyalty points notify them. If T goes by without the loyalty points being sorted out, notify the user that they will be charged M based on some conversion rate. This policy should be visible to customers when they use points to purchase stuff.
I don’t usually deal with microservices, and this might not be a good way of doing things, but here’s an idea:
To restate the problem, the system consists of three independent-but-communicating parts: the frontend, the order-management backend, and the loyalty-program backend. The frontend wants to make sure some state is saved in both the order-management backend and the loyalty-program backend.
One possible solution would be to implement some type of two-phase commit:
First, the frontend places a record in its own database with all the data. Call this the frontend record.
The frontend asks the order-management backend for a transaction ID, and passes it whatever data it would need to complete the action. The order-management backend stores this data in a staging area, associating with it a fresh transaction ID and returning that to the frontend.
The order-management transaction ID is stored as part of the frontend record.
The frontend asks the loyalty-program backend for a transaction ID, and passes it whatever data it would need to complete the action. The loyalty-program backend stores this data in a staging area, associating with it a fresh transaction ID and returning that to the frontend.
The loyalty-program transaction ID is stored as part of the frontend record.
The frontend tells the order-management backend to finalize the transaction associated with the transaction ID the frontend stored.
The frontend tells the loyalty-program backend to finalize the transaction associated with the transaction ID the frontend stored.
The frontend deletes its frontend record.
If this is implemented, the changes will not necessarily be atomic, but it will be eventually consistent. Let’s think of the places it could fail:
If it fails in the first step, no data will change.
If it fails in the second, third, fourth, or fifth, when the system comes back online it can scan through all frontend records, looking for records without an associated transaction ID (of either type). If it comes across any such record, it can replay beginning at step 2. (If there is a failure in step 3 or 5, there will be some abandoned records left in the backends, but it is never moved out of the staging area so it is OK.)
If it fails in the sixth, seventh, or eighth step, when the system comes back online it can look for all frontend records with both transaction IDs filled in. It can then query the backends to see the state of these transactions—committed or uncommitted. Depending on which have been committed, it can resume from the appropriate step.
I agree with what #Udi Dahan said. Just want to add to his answer.
I think you need to persist the request to the loyalty program so that if it fails it can be done at some other point. There are various ways to word/do this.
1) Make the loyalty program API failure recoverable. That is to say it can persist requests so that they do not get lost and can be recovered (re-executed) at some later point.
2) Execute the loyalty program requests asynchronously. That is to say, persist the request somewhere first then allow the service to read it from this persisted store. Only remove from the persisted store when successfully executed.
3) Do what Udi said, and place it on a good queue (pub/sub pattern to be exact). This usually requires that the subscriber do one of two things... either persist the request before removing from the queue (goto 1) --OR-- first borrow the request from the queue, then after successfully processing the request, have the request removed from the queue (this is my preference).
All three accomplish the same thing. They move the request to a persisted place where it can be worked on till successful completion. The request is never lost, and retried if necessary till a satisfactory state is reached.
I like to use the example of a relay race. Each service or piece of code must take hold and ownership of the request before allowing the previous piece of code to let go of it. Once it's handed off, the current owner must not lose the request till it gets processed or handed off to some other piece of code.
Even for distributed transactions you can get into "transaction in doubt status" if one of the participants crashes in the midst of the transaction. If you design the services as idempotent operation then life becomes a bit easier. One can write programs to fulfill business conditions without XA. Pat Helland has written excellent paper on this called "Life Beyond XA". Basically the approach is to make as minimum assumptions about remote entities as possible. He also illustrated an approach called Open Nested Transactions (http://www.cidrdb.org/cidr2013/Papers/CIDR13_Paper142.pdf) to model business processes. In this specific case, Purchase transaction would be top level flow and loyalty and order management will be next level flows. The trick is to crate granular services as idempotent services with compensation logic. So if any thing fails anywhere in the flow, individual services can compensate for it. So e.g. if order fails for some reason, loyalty can deduct the accrued point for that purchase.
Other approach is to model using eventual consistency using CALM or CRDTs. I've written a blog to highlight using CALM in real life - http://shripad-agashe.github.io/2015/08/Art-Of-Disorderly-Programming May be it will help you.
I am trying to use Mobile Data service in Bluemix and I come across the term last write wins. Can anyone explain what it is clearly?
And what are the other options apart from it
Last write wins is a strategy of deciding which data is most up-to-date when replication is used. Cassandra can be the example. It's simple & fast (uses timestamps) but has limited guarantees - it can cause lost updates / writes. The reason being is that time in computer systems isn't very accurate and nodes sometimes go down.
Check out CouchDB and MongoDB on how they perform consistency... MongoDB uses locks to achieve consistency while CouchDB uses eventual consistency. Mobile data is based on Cloudant (CouchDB under the covers) hence why it Mobile Data uses eventual consistency "last write wins".
Last write wins is basically used during synchronization of files for mobile applications through the File Sync plug-in.
File Sync is limited to use a "last write wins" policy when multiple applications are updating the same files. In "last write wins", the device's copy overwrites the copy stored by File Sync. The resulting behavior depends on whether you are running in automatic or manual mode."
You can visit below link for the reference:
https://www.ng.bluemix.net/docs/#services/mobiledata/index.html#filesync
My understanding of "last write wins" is like this:
Using Selenium i was the first user to book a tennis court on a website. However another user also managed to write to the same booking as we were within milliseconds of each other. The bookings have to be done at 7am. That gives us one whole second. I had managed to book the court and got the email notification that I had successfully booked this court. However, my colleague, who was slightly later (within the second) ended up being the last user to write (exit) to the booking (within the one second) and he also got a email notification. But much to my annoyance it was his name that appeared on the website as the user who had booked the court. His was the last write to the database and he wins - even though I beat him initially to the court booking. Once I had the final write there was probably a thousand of a second left enough for him to get the final write.
Ganz
My web app contains data gathered from an external API of which I do not have control. I'm limited to about 20,000 API requests per hour. I have about 250,000 items in my database. Each of these items is essentially a cached version. Consider that it takes 1 request to update the cache of 1 item. Obviously, it is not possible to have a perfectly up-to-date cache under these circumstances. So, what things should I be considering when developing a strategy for caching the data. These are the things that come to mind, but I'm hoping someone has some good ideas I haven't thought of.
time since item was created (less time means more important)
number of 'likes' a particular item has (could mean higher probability of being viewed)
time since last updated
A few more details: the items are photos. Every photo belongs to an event. Events that are currently occurring are more like to be viewed by client (therefore they should take priority). Though I only have 250K items in database now, that number increases rather rapidly (it will not be long until 1 million mark is reached, maybe 5 months).
Would http://instagram.com/developer/realtime/ be any use? It appears that Instagram is willing to POST to your server when there's new (and maybe updated?) images for you to check out. Would that do the trick?
Otherwise, I think your problem sounds much like the problem any search engine has—have you seen Wikipedia on crawler selection criteria? You're dealing with many of the problems faced by web crawlers: what to crawl, how often to crawl it, and how to avoid making too many requests to an individual site. You might also look at open-source crawlers (on the same page) for code and algorithms you might be able to study.
Anyway, to throw out some thoughts on standards for crawling:
Update the things that have changed often when updated. So, if an item hasn't changed in the last five updates, then maybe you could assume it won't change as often and update it less.
Create a score for each image, and update the ones with the highest scores. Or the lowest scores (depending on what kind of score you're using). This is a similar thought to what is used by LilyPond to typeset music. Some ways to create input for such a score:
A statistical model of the chance of an image being updated and needing to be recached.
An importance score for each image, using things like the recency of the image, or the currency of its event.
Update things that are being viewed frequently.
Update things that have many views.
Does time affect the probability that an image will be updated? You mentioned that newer images are more important, but what about the probability of changes on older ones? Slow down the frequency of checks of older images.
Allocate part of your requests to slowly updating everything, and split up other parts to process results from several different algorithms simultaneously. So, for example, have the following (numbers are for show/example only--I just pulled them out of a hat):
5,000 requests per hour churning through the complete contents of the database (provided they've not been updated since the last time that crawler came through)
2,500 requests processing new images (which you mentioned are more important)
2,500 requests processing images of current events
2,500 requests processing images that are in the top 15,000 most viewed (as long as there has been a change in the last 5 checks of that image, otherwise, check it on a decreasing schedule)
2,500 requests processing images that have been viewed at least
Total: 15,000 requests per hour.
How many (unique) photos / events are viewed on your site per hour? Those photos that are not viewed probably don't need to be updated often. Do you see any patterns in views for old events / phones? Old events might not be as popular so perhaps they don't have to be checked that often.
andyg0808 has good detailed information however it is important to know the patterns of your data usage before applying in practice.
At some point you will find that 20,000 API requests per hour will not be enough to update frequently viewed photos, which might lead you to different questions as well.
I have an asp.net-mvc website and people manage a list of projects. Based on some algorithm, I can tell if a project is out of date. When a user logs in, i want it to show the number of stale projects (similar to when i see a number of updates in the inbox).
The algorithm to calculate stale projects is kind of slow so if everytime a user logs in, i have to:
Run a query for all project where they are the owner
Run the IsStale() algorithm
Display the count where IsStale = true
My guess is that will be real slow. Also, on everything project write, i would have to recalculate the above to see if changed.
Another idea i had was to create a table and run a job everything minutes to calculate stale projects and store the latest count in this metrics table. Then just query that when users log in. The issue there is I still have to keep that table in sync and if it only recalcs once every minute, if people update projects, it won't change the value until after a minute.
Any idea for a fast, scalable way to support this inbox concept to alert users of number of items to review ??
The first step is always proper requirement analysis. Let's assume I'm a Project Manager. I log in to the system and it displays my only project as on time. A developer comes to my office an tells me there is a delay in his activity. I select the developer's activity and change its duration. The system still displays my project as on time, so I happily leave work.
How do you think I would feel if I receive a phone call at 3:00 AM from the client asking me for an explanation of why the project is no longer on time? Obviously, quite surprised, because the system didn't warn me in any way. Why did that happen? Because I had to wait 30 seconds (why not only 1 second?) for the next run of a scheduled job to update the project status.
That just can't be a solution. A warning must be sent immediately to the user, even if it takes 30 seconds to run the IsStale() process. Show the user a loading... image or anything else, but make sure the user has accurate data.
Now, regarding the implementation, nothing can be done to run away from the previous issue: you will have to run that process when something that affects some due date changes. However, what you can do is not unnecessarily run that process. For example, you mentioned that you could run it whenever the user logs in. What if 2 or more users log in and see the same project and don't change anything? It would be unnecessary to run the process twice.
Whatsmore, if you make sure the process is run when the user updates the project, you won't need to run the process at any other time. In conclusion, this schema has the following advantages and disadvantages compared to the "polling" solution:
Advantages
No scheduled job
No unneeded process runs (this is arguable because you could set a dirty flag on the project and only run it if it is true)
No unneeded queries of the dirty value
The user will always be informed of the current and real state of the project (which is by far, the most important item to address in any solution provided)
Disadvantages
If a user updates a project and then upates it again in a matter of seconds the process would be run twice (in the polling schema the process might not even be run once in that period, depending on the frequency it has been scheduled)
The user who updates the project will have to wait for the process to finish
Changing to how you implement the notification system in a similar way to StackOverflow, that's quite a different question. I guess you have a many-to-many relationship with users and projects. The simplest solution would be adding a single attribute to the relationship between those entities (the middle table):
Cardinalities: A user has many projects. A project has many users
That way when you run the process you should update each user's Has_pending_notifications with the new result. For example, if a user updates a project and it is no longer on time then you should set to true all users Has_pending_notifications field so that they're aware of the situation. Similarly, set it to false when the project is on time (I understand you just want to make sure the notifications are displayed when the project is no longer on time).
Taking StackOverflow's example, when a user reads a notification you should set the flag to false. Make sure you don't use timestamps to guess if a user has read a notification: logging in doesn't mean reading notifications.
Finally, if the notification itself is complex enough, you can move it away from the relationship between users and projects and go for something like this:
Cardinalities: A user has many projects. A project has many users. A user has many notifications. A notifications has one user. A project has many notifications. A notification has one project.
I hope something I've said has made sense, or give you some other better idea :)
You can do as follows:
To each user record add a datetime field sayng the last time the slow computation was done. Call it LastDate.
To each project add a boolean to say if it has to be listed. Call it: Selected
When you run the Slow procedure set you update the Selected fileds
Now when the user logs if LastDate is enough close to now you use the results of the last slow computation and just take all project with Selected true. Otherwise yourun again the slow computation.
The above procedure is optimal, becuase it re-compute the slow procedure ONLY IF ACTUALLY NEEDED, while running a procedure at fixed intervals of time...has the risk of wasting time because maybe the user will neber use the result of a computation.
Make a field "stale".
Run a SQL statement that updates stale=1 with all records where stale=0 AND (that algorithm returns true).
Then run a SQL statement that selects all records where stale=1.
The reason this will work fast is because SQL parsers, like PHP, shouldn't do the second half of the AND statement if the first half returns true, making it a very fast run through the whole list, checking all the records, trying to make them stale IF NOT already stale. If it's already stale, the algorithm won't be executed, saving you time. If it's not, the algorithm will be run to see if it's become stale, and then stale will be set to 1.
The second query then just returns all the stale records where stale=1.
You can do this:
In the database change the timestamp every time a project is accessed by the user.
When the user logs in, pull all their projects. Check the timestamp and compare it with with today's date, if it's older than n-days, add it to the stale list. I don't believe that comparing dates will result in any slow logic.
I think the fundamental questions need to be resolved before you think about databases and code. The primary of these is: "Why is IsStale() slow?"
From comments elsewhere it is clear that the concept that this is slow is non-negotiable. Is this computation out of your hands? Are the results resistant to caching? What level of change triggers the re-computation.
Having written scheduling systems in the past, there are two types of changes: those that can happen within the slack and those that cause cascading schedule changes. Likewise, there are two types of rebuilds: total and local. Total rebuilds are obvious; local rebuilds try to minimize "damage" to other scheduled resources.
Here is the crux of the matter: if you have total rebuild on every update, you could be looking at 30 minute lags from the time of the change to the time that the schedule is stable. (I'm basing this on my experience with an ERP system's rebuild time with a very complex workload).
If the reality of your system is that such tasks take 30 minutes, having a design goal of instant gratification for your users is contrary to the ground truth of the matter. However, you may be able to detect schedule inconsistency far faster than the rebuild. In that case you could show the user "schedule has been overrun, recomputing new end times" or something similar... but I suspect that if you have a lot of schedule changes being entered by different users at the same time the system would degrade into one continuous display of that notice. However, you at least gain the advantage that you could batch changes happening over a period of time for the next rebuild.
It is for this reason that most of the scheduling problems I have seen don't actually do real time re-computations. In the context of the ERP situation there is a schedule master who is responsible for the scheduling of the shop floor and any changes get funneled through them. The "master" schedule was regenerated prior to each shift (shifts were 12 hours, so twice a day) and during the shift delays were worked in via "local" modifications that did not shuffle the master schedule until the next 12 hour block.
In a much simpler situation (software design) the schedule was updated once a day in response to the day's progress reporting. Bad news was delivered during the next morning's scrum, along with the updated schedule.
Making a long story short, I'm thinking that perhaps this is an "unask the question" moment, where the assumption needs to be challenged. If the re-computation is large enough that continuous updates are impractical, then aligning expectations with reality is in order. Either the algorithm needs work (optimizing for local changes), the hardware farm needs expansion or the timing of expectations of "truth" needs to be recalibrated.
A more refined answer would frankly require more details than "just assume an expensive process" because the proper points of attack on that process are impossible to know.