If I am correct, crate (crate.io) is backed by Elasticsearch (Lucene). Weren't there a few articles a month ago that said that ES lost some writes under heavy load? Are there any other concerns?
You are right, Crate is backed by Elasticsearch. We think that the guys at elasticsearch are doing a great job on improving data consistency. A good read is http://www.elasticsearch.org/blog/resiliency-elasticsearch/ which gives a pretty good overview about efforts towards reliability. We at Crate are confident that this storage engine is safe to use as primary store. We also see, that issues regarding this area are getting actively worked on by the Lucene and Elasticsearch Community.
I am currently evaluating Crate.io as a primary datastore for work. As the above answer is vague and unspecific, maybe it's time for an update on this question here. There is a Dec.2016 keynote presentation out there from the Jepsen author Kyle Kingsbury on Youtube who investigated Crate.io for some problems with the resiliency in Elasticsearch. The first 8 minutes are introduction, the Crate.io part is from 23:50 till 31:10.
For those of you who don't want to watch the full video, here is a short summary.
First, the test setup. They set up databases and a random pattern of clients with random queries. Also, they voluntarily introduced problems for the databases, like network partitioning. Secondly, the results. According to Kingsbury, there are two issues with ES resiliency. Both of them carry on to Crate.io. Let's get to the details...
Dirty reads
The first one - ES #20031 - is that ES may cause dirty reads, divergence and lost updates if network partitions occur. As of now - December 2017 - this issue is still open. In my opinion, it is possible for the same problems to occur if a node is unresponsive for extremely heavy duty, like during extensive querying, reindexing or garbage collection.
Lost updates
According to Kingsbury, there is another problem ("Can promote stale binaries") with ES that cause updates to get completely lost when network partitioning occurs. It has been tagged as #20384 and there is kind of a fix which Kingsbury summarizes as "partial". So, ES may still cause lost data upon writing.
What does ES say?
On the official site of ES about resiliency, only one of the two problems - #20384 - is mentioned. It has been marked as solved in the version 5.0 release notes, although the official site says that there is only a partial fix.
What does Crate.io say?
On the Crate.io documentation on resiliency, there is a list of known problems with Crate.io resiliency. The ES bug #20384 is commented as partially fixed and still causing an open problem. The ES bug #20031 is not mentioned. However, there is a paragraph about an issue with networking partition which Crate.io marked as fixed - so the official page is kind of inconclusive here.
Conclusion
Kingsbury concluded in December 2016 that Crate.io should not be used as the primary data store. It could be used of course as a replication of your primary data to benefit from the time series database features that Crate.io offers. He also suggests that for machine data where 5% data loss is not a severe problem, Crate.io is a viable option as primary store.
It is my impression that some bugs Kingsbury reported may have been fixed but not all.
Related
I'm new to PipelineDB and have yet to even experience it at runtime (installation pending ...). But I'm reading over the documentation and I'm totally intrigued.
Apparently, PipelineDB is able to take set-based query representations and mechanically transform them into an incremental representation for efficiently processing the streams of deltas with storage limited as a function of the output of the continuous view.
Is it also supported to run the set-based query as a set-based query for priming a continuous view? It seems to me that upon creation of a Continuous View the initial data would be computed this traditional way. Also, since Continuous Views can be truncated, can they then be repopulated (from still-available source tables) without tearing down whatever dependent objects it has to allow a drop/create?
It seems to me that this feature would be critical in many practical scenarios. One easy example would be refreshing occasionally to reset the drift from rounding errors in, say, fractional averages.
Another example is if there were bug discovered and fixed in PipelineDB itself which had caused errors in the data. After the software is patched, the queries based on data still available ought to be rerun.
Continuous Views based fully on event streams with no permanent storage could not be rebuilt in that way. Not sure about if only part of the join sources are ephemeral.
I don't see these topics covered in the docs. Can you explain how these are or aren't a concern?
Thanks!
Jeff from PipelineDB here.
The main answer to your question is covered in the introduction section of the PipelineDB technical docs:
"PipelineDB can dramatically reduce the amount of information that needs to be persisted to disk because only the output of continuous queries is stored. Raw data is discarded once it has been read by the continuous queries that need to read it."
While continuous views only store the output of continuous queries, almost everybody who is using PipelineDB is storing their raw data somewhere cheap like S3. PipelineDB is meant to be the realtime analytics layer that powers things like realtime reporting applications and realtime monitoring & alerting systems, used almost always in conjunction with other systems for data infrastructure.
If you're interested in PipelineDB you might also want to check out the new realtime analytics API product we recently rolled out called Stride. The Stride API gives developers the benefit of continuous SQL queries, integrated storage, windowed queries, and other things like realtime webhooks, all without having to manage any underlying data infrastructure, all via a simple HTTP API.
If you have any additional technical questions you can always find our open-source users and dev team hanging out in our gitter chat channel.
Event sourcing and CQRS is great because it gets rids developers being stuck with one pre-modelled database which the developer has to work with for the lifetime of the application unless there is a big data migration project. CQRS and ES also has other benefits like scaling eventstore, audit log etc. that are already all over the internet.
But what are the disadvantages ?
Here are some disadvantages that I can think of after researching and writing small demo apps
Complex: Some people say ES is complex. But I'd say having a complex application is better than a complex database model on which you can only run very restricted queries using a query language (multiple joins, indexes etc). I mean some programming languages like Scala have very rich collection library that is very flexible to produce some seriously complex aggregations and also there is Apache Spark which makes it easy query distributed collections. But databases will always be restricted to it's query language capabilities and distributing databases are harder then distributed application code (Just deploy another instance on another machine!).
High disk space usage: Event store might end up using a lot of disk space to store events. But we can schedule a clean up every few weeks and creating snapshot and may be we can store historical events locally on an external HD just incase we need old events in the future ?
High memory usage: State of every domain object is stored in memory which might increase RAM usage and we all how expensive RAM is. BIG PROBLEM!! because I'm poor! any solution to this ? May be use Sqlite instead of storing state in memory ? Am I making things more complex by introducing multiple Sqlite instances in my application ?
Longer bootup time: On failure or software upgrade bootup is slow depending on the number of events. But we can use snapshots to solve this ?
Eventual consistency: Problem for some applications. Imagine if Facebook used Event sourcing with CQRS for storing posts and considering how busy facebook's system is and if I posted a post I would see my fb post the next day :)
Serialized events in Event store: Event stores store events as serialized objects which means we can't query the content of events in the event store which is discouraged anyway. And we won't be able to add another attribute to the event in the future. Solution would be to store events as JSON objects instead of Serialized events ? But is that a good idea ? Or add more Events to support the change to orignal event object ?
Can someone please comment on the disadvantages I brought up here and correct me if I am wrong and suggest any other I may have missed out ?
Here is my take on this.
CQRS + ES can make things a lot simpler in complex software systems by having rich domain objects, simple data models, history tracking, more visibility into concurrency problems, scalability and much more. It does require a different way thinking about the systems so it could be difficult to find qualified developers. But CQRS makes it simpler to separate responsibilities across developers. For example, a junior developer can work purely with the read side without having to touch business logic.
Copies of data will require more disk space for sure. But storage is relatively cheap these days. It may require the IT support team to do more backups and planning how to restore the system in a case in things go wrong. However, server virtualization these days makes it a more streamlined workflow. Also, it is much easier to create redundancy in the system without a monolithic database.
I do not consider higher memory usage a problem. Business object hydration should be done on demand. Objects should not keep references to events that have already been persisted. And event hydration should happen only when persisting data. On the read side you do not have Entity -> DTO -> ViewModel conversions that usually happened in tiered systems, and you would not have any kind of object change tracking that full featured ORMs usually do. Most systems perform significantly more reads than writes.
Longer boot up time can be a slight problem if you are using multiple heterogeneous databases due to initialization of various data contexts. However, if you are using something simple like ADO .NET to interact with the event store and a micro-ORM for the read side, the system will "cold start" faster than any full featured ORM. The important thing here is not to over-complicate how you access the data. That is actually a problem CQRS is supposed to solve. And as I said before, the read side should be modeled for the views and not have any overhead of re-mapping data.
Two-phase commit can work well for systems that do not need to scale for thousands of users in my experience. You would need to choose databases that would work well with the distributed transaction coordinator. PostgreSQL can work well for read and write separate models, for example. If the system needs to scale for a high number of concurrent users, it would have to be designed with eventual consistency in mind. There are cases where you would have aggregate roots or context boundaries that do not use CQRS to avoid eventual consistency. It makes sense for non-collaborative parts of the domain.
You can query events in serialized a format like JSON or XML, if you choose the right database for the event store. And that should be only done for purposes of analytics. Nothing inside the system should query event store by anything other than the aggregate root id and the event type. That data would be indexed and live outside the serialized event.
Just to comment on point 5. I've been told that Facebook does use ES with Eventual Consistency, which is why you can sometimes see a post disappear and reappear after you've posted it.
Usually the read-model your browser is accessing is located 'close' to you, but after you make a post the SPA switches over to a read-model that is close to your write-model. The close proximity between the write-model (events) and the read-model mean you get to see your own post.
However, 15 minutes later your SPA switches back to the first, closer, read-model. If the event containing your post hasn't yet propagated to that read-model you'll see your own post disappear only to reappear sometime later.
I know it's been almost 3 years since this question was asked, but still this article may be useful for someone. Key points are
Scaling with snapshots
Visibility of data
Schema changing
Dealing with complex domains
Need to explain it to most new team members
Event sourcing and CQRS is great because it gets rids developers being stuck with one pre-modeled database which the developer has to work with for the lifetime of the application unless there is a big data migration project.
This is a big misconception. The relational databases were invented exactly for the evolution of the model (thanks to simple two-dimensional tables as opposed to pre-defined hierarchical structures). With views and procedures ensuring the encapsulation of data access, the logical and physical model can evolve independently. This is also why SQL defines DDL and DML in the same language. Some RDBMS also allow all those evolutions to be versioned and deployed online (continuous delivery) as Oracle Edition Based Redefinition.
Big data structures are predefined and can be read only with the code developed for this structure. Ok when consumed immediately but you will have hard time to read it 10 years later without the exact version, and language compiler or interpreter.
I hope to not be late to try to give an answer. In these months I've done a lot of research on that argument with the goal of implementing a production-grade solution for some parts of my architecture where ES can make sense
Complex: Actually, it should not be complex, its mission is to be deadly simple. How? pushing all the complexity from business logic code to infrastructure code. The data access should be done by frameworks that are not enough mature yet. Still, there is no clear winner in the ES/CQRS race, maybe because is still a niche/hipster approach (?) So some team is rolling its own solution or adopting some ready-made technology such as Axon
High disk space usage: I would say more, I would say * potentially infinite* Disk Usage. But if you go towards ES, you also have a very good reason to tolerate this apparent drawback. Let's give some of them:
Audit Logs : The datastore is an event log, we already know it. Financial apps or every mission/safety critical could need a centralized audit log that enables to state Who made What in Which moment. ES provides this capability of the box...you can also decorate your event entries with some business meaningful metadata (eg. a transaction Id correlated with some API consumer identity, A severity level of the operation...)
High Concurrency: there are systems where logical resource states are mutated by many clients in a concurrent way. These are games, IoT platforms, and so on. Logging events instead of change a state representation could be a smart way to provide a total order of events. The other way is to delegate to DB the synchronization stuff. But this is not what you want if you're into ES
Analytics Let's say you have a lot of data with a lot of business value, but you still don't know which. For years we extracted knowledge from applications information by translating data organization with different information models (OLAP cubes). The event store provides something similar out of the box again. Event logs is the rawest form of representation of information And you can have many ways to process them, in batch or reacting to events stored
High memory usage: I think it should be the same once you have built your projection
Longer bootup time: If the read side caches its projections and "remembers" the last update event, it should not re-apply the entire event sequence. Snapshots are mitigation but if you do a lot of snapshots maybe you made a bad choice with ES. I think that this problem is minor in microservices ecosystems, where the boot time can be masked without service interruption. In fact, you get the most out of ES/CQRS when you apply it so microservices
Eventual consistency: Blame CAP theorem for this, not ES. Many non ES/CQRS have to deal with this, but there are a lot of scenarios where it is not a real problem. These are the scenarios where ES fits well. And you can mix ES and non ES services into the same platform
Serialized events in Event store: if it's important to have a non-serialized event representation, you could use a document-oriented DB, but if you do this to make queries over events payload, you are missing the point of ES/CQRS. ES means to move all data manipulation from the DB side to the application tier, where every piece changes fastly, and all are stateless. This enhances scalability and fault tolerance and provides means to shape the organization of your team, doing things like let the frontend guy/girl write his/her BFF in javascript easily.
I hope to put into practices this principles with good results and draw on the benefits of this exciting approach
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 6 years ago.
Improve this question
I would like to know what advantages there are to using EventStore (http://geteventstore.com) over implementing event sourcing yourself in a MongoDb.
The reason I ask, is that our company has a number of people that work with MongoDb daily. They don't work with Event Sourcing though. While they are not completely in the dark about the subject, they aren't about to start implementing it anywhere either.
I am about to start a project, that is perfectly suited for Event Sourcing. There are about 16 very well defined events, and about 7 well defined projections. I say "about" because I know there will be demand for more projections and events once they see the product in use.
The approach is going to be API first, with a REST Api that other parts of our organisation are going to consume.
While I have read a lot about Event Sourcing the way Greg Young defines it, I have never actually implemented an Event Sourcing solution.
This is a green field project. No technology restrictions since we are going to expose everything as a REST interface. So if anyone has working experience with EvenStore or Event Sourcing with MongoDb please enlighten me.
Also an almost totally non related question about Event Sourcing:
Do you ever query the event store directly? Or would you always create new projections and replay event to populate those projections?
Disclaimer I am Greg Young (if you cant read my name :))
I am going to answer this question though I believe it will likely get deleted anyways. This question alone for me is a bit odd, but the answers are fairly bizarre. I won't take the time to answer each reply individually but will instead put all of my comments in this reply.
1) There is a comment that we only run on a custom version of mono which is a detail but... This is not the case (and has not been for over a year). We were waiting on critical patches we made to mono (as example threadpool.c to hit their master). This has happened.
2) EventStore is 3-clause BSD licensed. Not sure how you could claim we are not Open Source. We also have a company behind it and provide commercial support.
3) Someone mentioned us going on to version 3 in Sept. Version 1 was released 2 years ago. Version 2 added Clustering (obviously some breaking changes vs single node). Version 3 is adding a ton of stuff including ability to have competing consumers. Very little has changed in terms of the actual client protocol over this time (especially for those using the HTTP API).
What is really disturbing for me in the recommendations however is that they don't seem to understand what they are comparing. It would be roughly the equivalent of me saying "Which should I use neo4j or leveldb?". You could build yourself a graph database on top of leveldb but that would be quite a bit of work.
Mongo in this case would be a storage engine on the event store the OP would have to write him/herself. The writing of a production quality event store is a non-trivial exercise on top of a storage engine if you want to have even the most basic operations.
I wrote this in response to the mailing list equivalent of this question:
How will you do the following with Mongo?:
Write and read events to/from streams with ordering/optimistic concurrency/etc
Then:
Your projections don't want to read from streams in the same way they were written, projections are normally interested in event types and want all events of type T regardless of stream written to and in proper order.
You probably also want for instance the ability to switch live from pushed event notifications to handling pulled information (eg polling) etc.
It would make more sense if Kafka, datomic, and Event Store were being compared.
Seeing as the other replies don't talk about the tooling or benefits in EventStore and only refer to the benefits of MongoDB I'll chime in. But note that my experience is limited.
I'll start with the cons...
There are a lot of check-ins which can lead to deciding which version you are going to actively support yourself. While the team has been solidifying their releases, that they have arrived at version 3 not even 18 months after being released should be an indicator that you have to pull up the version you are supporting for another more recent version (which can also impact the platform you choose to deploy to).
It's not going to easily work on every platform (especially if you're trying to move to a cloud environment or a docker based lxc container). Some of this is due to the community surrounding other DBs such as Mongo. But the team seems to have been working their butts off on read/write performance while maintaining cross platform stability. As time presses on I've found that you don't want to deviate too far from a bare-metal OS implementation which this day in age is not attractive.
Uses a special version of Mono. Finding support for older versions of Mono only serve to make the process more of a root canal.
To make the most of performance of EventStore you really need to think about your architecture. EventStore outputs to flat files and event data can grow pretty quickly. What's the fail rate of the disks are you persisting your data to. How are things compressed? archived? etc. You have a lot of control and the control is geared towards storing your data as events. However, while I'm sure Greg Young himself could quote me to my grave the features that optimize and save your disks in the long term, I'll more than likely find a mature Mongo community that has had experience running into similar cases.
And the Pros...
RESTful - It's AtomPub. Is your stream not specific enough? Create another and do http gets till your hearts content. Concerned about routing do do an http forward. Concerned about security put an http proxy in front. Simple!
You have a nice suite of tools and UI for testing out and building your projections as your events start to generate new data (eg. use chrome browser as a way to debug your projections... ya they're written with java script)
Read performance - Since the application outputs to a flat file you can get kernel level caching and expose them via http in the drop of a hat. Also indexes are across your streams for querying projections against larger data sets (but I really get the feeling index performance will creep up on you over time).
I personally would not use this for a core / mission critical / or growing application! However, if you have a side case for keeping your evented environment interesting then I'd give it go! I personally have to stick to Mongo for now.
On the postgreSQL's wiki, on the "Replication, Clustering, and Connection Pooling" page ( http://wiki.postgresql.org/wiki/Replication,_Clustering,_and_Connection_Pooling) , it shows the following example on replication's requirements:
"Your users take a local copy of the database with them on laptops when they leave the office, make changes while they are away, and need to merge those with the main database when they return. Here you'd want an asynchronous, lazy replication approach, and will be forced to consider how to handle conflicts in cases where the same record has been modified both on the master server and on a local copy"
And that's pretty much my case. But, unfortunatelly, on the same page, it says: "(...) A great source for this background is in the Postgres-R Terms and Definitions for Database Replication. The main theoretical topic it doesn't mention is how to resolve conflict resolution in lazy replication cases like the laptop situation, which involves voting and similar schemes."
What I want to know, is where can I find material on how to resolve this kind of situation, and wich would be the best way to do this on PostgreSQL.
I will have to check into RubyRep but it seems like Bucardo might be a more widely supported option.
Gabriel Weinberg has an EXCELLENT tutorial on his site for how he uses Bucardo. The guy runs his own search engine called DuckDuckGo and there are quite a few tips and tricks that are optimized for his use cases.
http://www.gabrielweinberg.com/blog/2011/05/replicating-postgresql-with-bucardo.html
Just answering my own question, if anyone ever finds it: I'm using Rubyrep http://www.rubyrep.org/ and it's working.
I want to build a web app similar to Reddit.com, where you have multy level of comments, lots of reads and writes. I was wondering if nosql and mongoDB in particular is the right tool for this?
Comments -- it's really thing for nosql database, no doubt. You avoiding multiple joins to itself. And it's means that your system can scale out!
With mongodb you can store all hierarchy within one document. Some peoples can say that here will be problems with atomic updates, but i guess that it's not a problem because of you can load and save back entire comments tree. In any way you can easy redesign your system later to support atomic updates and avoid issues with concurrency.
Reddit itself uses Cassandra. If you want something "similar to reddit.com," maybe you should look at their source -- https://github.com/reddit/reddit/wiki.
Here's what David King (ketralnis) said earlier this year about the Cassandra 0.7 release: "Running any large website is a constant race between scaling your user base and scaling your infrastructure to support it. Our traffic more than tripled this year, and the transparent scalability afforded to us by Apache Cassandra is in large part what allowed us to do it on our limited resources. Cassandra v0.7 represents the real-life operations lessons learned from installations like ours and provides further features like column expiration that allow us to scale even more of our infrastructure."
However, Rick Branson notes that Reddit doesn't take full advantage of Cassandra's features, so if you were to start from scratch, you'd want to do some things differently.