I'm learning Docker Swarm mode and I managed to create a Swarm locally with a web application and a PostgreSQL database. I can scale them and I see Swarm creating replicas.
I think I understand how Docker Swarm can load balance regular web servers, but how does it deal out of the box with database containers?
Outside of the Swarm context, usually databases have their own ways to deal with replication, in the form of plugins or extended products like MySQL cluster. Other databases like Cassandra have replication built directly into their product.
On a Swarm context, do we still need to rely on those database plugins and features?
What is the expected pattern to handle data consistency between replicas of a database container?
I know it's a very open-ended question, but Docker's documentation is very open-ended too and I can't seem to find anything specific to this.
How does it deal out of the box with database containers?
It doesn't.
There is a pretty good description of Swarm services here: How services work (emphasis mine)
When you deploy the service to the swarm, the swarm manager accepts your service definition as the desired state for the service. Then it schedules the service on nodes in the swarm as one or more replica tasks.
Swarm has no idea what's inside the task, all it knows is how many instances of it there are, whether those instances are passing their health checks, and if there are enough of them to satisfy the task definition you gave it. The word overlap between this and database replicas is a little unfortunate, but they are different concepts.
What is the expected pattern to handle data consistency between replicas of a database container?
Setting up data replication is on you. These are probably as good a place to start as any
How to Set Up PostgreSQL for High Availability and Replication with Hot Standby
PostgreSQL Replication Example
Docker swarm currently scales well for the stateless applications. For database replication, you have to rely on every database's own replication mechanism. Swarm could not manage the datatbase replication. The volume or file system level replication could provide the protection for a single instance database, but are not aware of database replication/cluster.
For databases such as PostgreSQL, the additional works are required. There are a few options:
Use host's local directory. You will need to create one service for every replica, and use constraint to schedule the container to one specific host. You will also need custom postgresql docker image to set up the postgresql replication among replicas. While, when one node goes down, one PostgreSQL replica will go down. You will need to work to bring up another replica. See crunchydata's example.
Use the volume plugin, such as flocker, REX-Ray. You will still need to create one service for every replica, and bind one volume to one service. You need to create all services in the same overlay network and configure the PostgreSQL replicas to talk with each other via the dns name (the docker service name of the replica). You will still need to set up the postgresql replication among replicas.
Related
In a scenario where there are multiple containers running multiple instances of a database, when data in one instance is changed, how that change is applied to other instances so the data consistency preserved? it should be taken into consideration that containers may have bee placed in different nodes, so they can't share a common volume.
Deploying database to Kubernetes in high-availability mode is a quite complex task.
To save a lot of time, please consider using Kubernetes operators. This way you can automate most of complex manual operations. Kubernetes operators allow you to describe desired configuration of an application (or database) in declarative way.
Kubernetes operator in general is a set of additional objects (pods, services, configmaps, secrets, RBAC, custom resources definitions, PV, PVC, ...) responsible to create and maintain complex application lifecycle.
Usually Kubernetes operators related to database applications take care about deploying database cluster of the desired size to Kubernetes and creating all required resources. Also they keep desired number of replicas, implement backup/restore procedures, configure data replication, forward client traffic to a healthy database pod and more.
Most of well-known databases have ready-to-use operators already. You can find some resources below:
Mysql Resources:
Introducing the Oracle MySQL Operator for Kubernetes
Oracle MySQL Operator on GitHub
Getting started with the MySQL Operator for Kubernetes
Percona Resources:
Percona Kubernetes Operator for Percona XtraDB Cluster
MySQL Operator
MySQL Operator on Github
PostgreSQL resources:
Crunchy Data PostgreSQL Operator
Postgres Operator
MongoDB resources:
Install the MongoDB Enterprise Kubernetes Operator
In case you are planning to use different type of database, please check for available operator from DB vendor searching "DBname kubernetes operator"
If you are going with normal deployment with an application that just uses folder to access data it is advisable to use Persistent volumes, it could be either be nfs or several other options mentioned in the documentation where you mount the same volume inside all different pods having access to same data.
And if you are trying to host a database which is an application with state, it is advisable to use a statefulset instead of normal deployment when coming to applications like mongodb, mysql, etc.
Here is a how the statefulset works. But there are several online resources that gives you the differences between using a deployment vs statefulset for databases. But however, before statefulsets were introduced into k8s many of our database applications were deployed using PVC, it also has lot of overhead.
Hope this is helpful.
Assuming we are using a micro service architecture for a product and we decide to use 'Database per service' model, and deploy in cloud servers by provider like AWS.
It is convenient to have databases running as a container for development and test environments.
But can same be implemented for Production environment! If so, how safe it would be?
Or is it proper to go with cloud solution as AWS RDS-DB instead!!
This blog post lists some reasons why you should not run production databases in containers. It also references another blog post describing problems with updating docker and unstable storage drivers.
The main points here for me boil down to this:
Dodgy storage drivers. This may be less of a problem when you write your database state to the host system but Docker for example explicitly encourages users to use volumes for exactly that (see the docs: Citation: "Volumes are the best way to persist data in Docker"). It may just work fine under normal circumstances, but what about the edge-cases like power-failures or read-errors for example?
Managing databases in production is hard. Many companies employ full-time DBAs to ensure smooth operation of production databases. The devops paradigm (every dev creates a plethora of DB servers in containers) makes it nearly impossible for a DBA to do his job. That is if the DBA even has access to these DBs.
In conclusion: Containers are fine for certain tasks and a bad idea for others. Running production databases in containers is one of those bad ideas.
We containerise our db in production (on-premises enterprise application). Many do. It's perfectly stable and the deployment is much simplified. Of course our db is not under stress; we're dealing with hundreds of concurrent users, not tens of thousands. We just make sure that the container has enough RAM and is monitored well.
If we did need to dedicate an entire VM to the db alone, then yes I would skip docker.
According to link below, it is not a good idea to use database container in Production.
But as I have experienced; if you isolate your container from your app and update your container regularly and also manage networking stuff, there seems to be no problem.
Link: https://www.quora.com/Is-it-not-advisable-to-use-database-in-Docker-container
As you are using Database Per Service Model for Microservice, in Production perfect solution can be AWS RDS instance for database, Now you have 2 approaches :
You can create single RDS Instance and can have different databases for different services on same RDS insatnce, it will save cost a lot but you need to take care of database connections and load you will be having on database based on that you have to choose RDS instance type like 4xlarge etc, better the instance type more connection it will provide and more database load it can handle effectively.
Second solution can be creating several RDS instance and number of RDS instance will be equivalent to your microservice count as each service will be using one RDS instance for its database independently, this is not the effective solution, it will incur lot of cost and this solution will under utilize AWS RDS instances.
What is the best practice to deploy database in microservices architecture, more precisely in distributed environment, such as docker swarm? Microservices principles states each service should be stateless to enable scaling. As database obviously has a state, should it live at fixed position outside of cluster, deployed and configured before the cluster is initialized?
I'm confused, because all docker compose examples includes database container in the service definition. But things aren't that simple. Often the database needs a lot of configuration before it's ready to use. Also, docker sucks at coordinating the service starting order.
If it's really a good practice to deploy the database alongside with services to docker swarm, how to ensure consistency and persistence of cricial data?
This is a good question and one I think a lot of people are still thinking through as far as best practices are concerned. The answer really depends on your needs. There are several ways to crack this nut but these are the two I'm using right now:
Running the database in the typical manner on dedicated machine(s) with replication, etc
I am currently experimenting with running the database as a service on a Docker Swarm cluster with the data persisted across the cluster with GlusterFS
I have three machines in the cluster labeled as database machines
These database machines all run a GlusterFS container providing the GlusterFS capabilities
When the database service is started I map the GlusterFS share into the container and specify that the service should only run on a machine labeled as a database node. With this setup it doesn't matter which node the database service starts on and if a machine fails the database service is automatically migrated to another node labeled as a database node. The GlusterFS replication of data ensures the integrity of the persisted data.
As mentioned, it is my understanding that there is still a lot of experimentation going on with this and 'best practices' are not entirely established. Those best practices will ultimately depend on your needs and risk tolerances.
With the introduction of the new 'swarm mode' with Docker 1.12, we've been trying to migrate our application on containers and make use of the swarm mode's orchestration & clusters.
Our application requires some initial database scripts to be run for it to start.
We're not packaging the database inside our dockerized application so that it could follow a stateless microservice architecture and multiple containers would eventually talk to a single (at the moment) database instance.
While creating the service, we cannot use --replicas with the create service command as multiple instances would try and create tables on a single database and fail. Although our scripts would check if the database has been set-up and skip the creation but since all containers start simultaneously, it could not be used.
We couldn't find any wait-for kind of mechanism that we could leverage with dockers for this issue. It would have been good if we could only start the second container when the first one had created the database (and exposed the ports) but how can we configure inter-container communication for this?
Alternatively, can tools like flywaydb help in some way?
How should this be used in production?
From the Flyway FAQ:
Can multiple nodes migrate in parallel?
Yes! Flyway uses the locking technology of your database to coordinate multiple nodes. This ensures that even if even multiple instances of your application attempt to migrate the database at the same time, it still works. Cluster configurations are fully supported.
There is no easy way to coordinate this among containers. It basically requires a distributed lock solution. The first container that gets the lock could create db, while, other containers that not get the lock need to wait.
In AWS, you could leverage DynamoDB for it. DynamoDB supports conditional update. The container first tries to create the lock key in DynamoDB with "attribute_not_exists(yourKey)". The first creation will succeed and other creations will be rejected. The first container needs to create another key in DynamoDB to indicate the db is ready. Other containers simply waits till the ready key is created.
Or you could do it in your service deployment script. The script could create the service with 1 replica. Then keep checking if db is created. If yes, scale the service, such as docker service update yourservicce --replicas 5.
I'm researching cloud services to host an e-commerce site. And I'm trying to understand some basics on how they are able to scale things.
From what I can gather from AWS, Rackspace, etc documentation:
Setup 1:
You can get an instance of a webserver (AWS - EC2, Rackspace - Cloud Server) up. Then you can grow that instance to have more resources or make replicas of that instance to handle more traffic. And it seems like you can install a database local to these instances.
Setup 2:
You can have instance(s) of a webserver (AWS - EC2, Rackspace - Cloud Server) up. You can also have instance(s) of a database (AWS - RDS, Rackspace - Cloud Database) up. So the webserver instances can communicate with the database instances through a single access point.
When I use the term instances, I'm just thinking of replicas that can be access through a single access point and data is synchronized across each replica in the background. This could be the wrong mental image, but it's the best I got right now.
I can understand how setup 2 can be scalable. Webserver instances don't change at all since it's just the source code. So all the http requests are distributed to the different webserver instances and is load balanced. And the data queries have a single access point and are then distributed to the different database instances and is load balanced and all the data writes are sync'd between all database instances that is transparent to the application/webserver instance(s).
But for setup 1, where there is a database setup locally within each webserver instance, how is the data able to be synchronized across the other databases local to the other web server instances? Since the instances of each webserver can't talk to each other, how can you spin up multiple instances to scale the app? Is this setup mainly for sites with static content where the data inside the database is not getting changed? So with an e-commerce site where orders are written to the database, this architecture will just not be feasible? Or is there some way to get each webserver instance to update their local database to some master copy?
Sorry for such a simple question. I'm guessing the documentation doesn't say it plainly because it's so simple or I just wasn't able to find the correct document/page.
Thank you for your time!
Update:
Moved question to here:
https://webmasters.stackexchange.com/questions/32273/cloud-architecture
We have one server setup to be the application server, and our database installed across a cluster of separate machines on AWS in the same availability zone (initially three but scalable). The way we set it up is with a "k-safe" replication. This is scalable as the data is distributed across the machines, and duplicated such that one machine could disappear entirely and the site continues to function. THis also allows queries to be distributed.
(Another configuration option was to duplicate all the data on each of the database machines)
Relating to setup #1, you're right, if you duplicate the entire database on each machine with load balancing, you need to worry about replicating the data between the nodes, this will be complex and will take a toll on performance, or you'll need to sacrifice consistency, or synchronize everything to a single big database and then you lose the effect of clustering. Also keep in mind that when throughput increases, adding an additional server is a manual operation that can take hours, so you can't respond to throughput on-demand.
Relating to setup #2, here scaling the application is easy and the cloud providers do that for you automatically, but the database will become the bottleneck, as you are aware. If the cloud provider scales up your application and all those application instances talk to the same database, you'll get more throughput for the application, but the database will quickly run out of capacity. It has been suggested to solve this by setting up a MySQL cluster on the cloud, which is a valid option but keep in mind that if throughput suddenly increases you will need to reconfigure the MySQL cluster which is complex, you won't have auto scaling for your data.
Another way to do this is a cloud database as a service, there are several options on both the Amazon and RackSpace clouds. You mentioned RDS but it has the same issue because in the end it's limited to one database instance with no auto-scaling. Another MySQL database service is Xeround, which spreads the load over several database nodes, and there is a load balancer that manages the connection between those nodes and synchronizes the data between the partitions automatically. There is a single access point and a round-robin DNS that sends the requests to up to thousands of database nodes. So this might answer your need for a single access point and scalability of the database, without needing to setup a cluster or change it every time there is a scale operation.