Recently, the team has been discussing improvements around Hibernate (ORM) usage within cloud based apps and microservices. In particular the fundamental assumption that things will break regularly on these platforms and that services should be resilient to failures.
In microservices or cloud architectures, services are started in different orders (usually beyond your control). It is possible for the app using Hibernate ORM to be started before the database. At the moment, Hibernate ORM does not like that and will explicitly fail (exception) if it can’t connect to the database.
Another related concern is to consider what is happening if the database is stopped for a while after the app running Hibernate ORM has started and resume working shortly after.
Solution 1: Hibernate waits and retries at boot time
Some users have asked us to delay and retry the connection process in case the database is not present at boot time. That would work and solve the bootstrap problem. It would not solve the database gone while running the app but here at least you have your transaction and the error propagation mechanism covering you. Plus at development time, the boot time problem gone would be quite nice already.
I understand that this is probably a quick win to implement this, but better be sure of the problem before adding that feature. It feels to me that Hibernate ORM bootstrap is not the ideal area to fix that problem. But at the end of the day if it helps enough, it would be worth it.
We are exploring that option and considering alternatives and that’s where we need your feedback.
Wait and retry vs platform notification
In this blog post, I mention the wait and retry approach. It can be replaced by a notification from the cloud platform when a service is up / down.
This avoids the regular polling process at the cost of having to rely on various integrations from various cloud platforms.
Solution 1.b: The connection pool waits and retries
It probably would be better if the connection pool Hibernate ORM uses, implements that logic but there are more than one connection pool Hibernate supports. That’s a minor variation on solution 1.
Solution 2: Hibernate boots in non-functioning mode
If Hibernate ORM cannot connect to the datasbase, it continues its bootstrap process.
EntityManager is asking for a connection while the database is still unavailable, a well defined exception is raised.
To not flood the system, a wait and retry system for connection checking would be in place to only try a few times even when lots of
EntityManager are requested.
There are some subtle difficulties here on concurrency and on the fact that we use info from the bootstrap connection to configure Hibernate ORM. The most visible option guessed from the connection is the dialect to use. On the other hand, stopping the app boot process while waiting and retrying like solution 1 proposes is probably not without its challenges.
The exception raised by Hibernate ORM upon DB inaccessibility needs to be treated properly by the application (framework) being used. Like a global try catch that moves the application in degraded mode or propagating the execution error to the client (e.g. HTTP error 500). It might even be helpful if Hibernate ORM was exposing the not ready status via an explicit API.
This could be tied to a health check from the cloud platform.
The application would report the not ready but trying status via a
/health endpoint that the orchestrator would use.
On database connection breaking
There are many reasons for failing to connect to a database:
Should the system go into the wait and retry mode for cases 3, 4, 5? Or should it refuse to deploy?
Solution 3: the smart app (framework)
Another solution is for the app to have a smart overall bootstrap logic. It tries to eagerly start but if a Hibernate ORM connection error occurs, only the inbound request framework is started. It will regularly try and boot and in the mean time return HTTP 500 errors or similar.
This requires an app framework that could handle that. It embeds circuit breaker logic in the app and can better react to specific errors. I wonder how common such frameworks are though.
This is in spirit the same solution as solution 2 except it is handled at the higher level of the app (framework) vs Hibernate ORM.
Solution 4: the cloud / MSA platform restart the apps
An arguably better solution would be for the cloud platform to handle these cases and restart apps that fail to deploy in these situations. It likely requires some kind of service dependency management and a bit of smartness from the cloud infra. The infrastructure would upon specific error code thrown at boot time, trigger a wait and retry deployment logic. There is also a risk of a dependency circularity leading to a never starting system.
I guess not all cloud infra offer this and we would need an alternative solution. OpenShift let’s you express dependencies to make sure a given service is started before another. The user would have to declare that dependency of course.
Solution 5: proxy!
Another solution is to put proxies either before the app inbound requests and/or between the app and the database. Proxy is the silver bullet that lots for cloud platforms uses to solve world hunger in the digital universe.
How many proxies and routing logic does it take to serve a "Hello world!" in the cloud?
This approach has the advantage of not needing customized apps or libraries. The inconvenience is more intermediary points between your client and the app or data.
If the proxy is before the application, then it needs a health check or a feedback from the boot system to wait and retry the re-deployment of the application on a regular basis. I’m again not certain cloud infrastructures offer all of this infrastructure.
If the proxy is between Hibernate ORM and the database (like HAProxy for MySQL), you’re still facing some timeout exception on the JDBC side. Which means the application will fail to boot. But at least, the proxy could implement the wait and retry logic.
My questions to you
Do you have any input on this subject:
what’s your opinion?
what’s your experience when deploying cloud apps?
any alternative solution you have in mind?
any resource you found interesting covering this subject?
would you benefit from solution 1?
would you benefit from solution 2?
Any feedback to help us think this problem further is what we need :)