Skip to main content

Azure CosmosDB Request Unit considerations

So I have been doing a deep dive into Azure CosmosDB for a project. One of my objectives was to learn its performance characteristics when uploading a few million records an hour, and doing a cost analysis of CosmosDB vs other technologies already in use.  I put together an app that's able to push a few hundred records in at an acceptable rate.  So I do what I inevitably always do. From my laptop on hotel internet, I "dial it up to 11, and see what brakes", what could go wrong?

Well, it worked, and as expected, throttled when it hit the RU (Request Unit) limit.  I increased the RU allocation, and it pushed more documents per second.  Cool? NO, I was pushing 10,000+ RU/s and only getting a few hundred documents a second.  So I did what I should have done to start with, ran the sample from a VM in the same datacenter as my DB.  BINGO, it easily surpassed 1,000 documents per second. But for the same RU/s?

What's going on? Well after a bit of testing, I suspect that the RU calculation is primarily, if not exclusively, a function of time between when you open that socket and you get the "200 OK" back.  That includes how long it takes you to send the data.  Meaning that 10 MBs of documents sent at 1MB/s will be many times more expensive than the same documents sent at 1GB/s.

TL;DR : CosmosDB Request units are highly susceptible to connection latency.  So under load, even the latency between an Azure data center in Europe and the US will increase your RU/s used by a small amount.

So what do we do about this?  It's another factor you need to consider when selecting your replica data centers.  And you may get more than just lower latency on writes when using a "follow the sun" strategy with your master node.
Labels:

Popular posts from this blog

Getting Azure AD JWT via Postman

In my last post, I reviewed how to use Azure AD for service to service authentication.  But sometimes, you may want to test your API directly.  So let's review how to get the token via Postman. We are going to be using the OAuth2 endpoint, and going for a "client_credentials" grant type.  So you are going to need to know a few things from your Azure portal. Application Id : This is used for the Client Id. Application key : This is used for the Client Secret.  You have to generate this from the portal, and it will only give it back to you once.  If you lose it, you will have to generate a new one. Application ID URI : we will use this as our resource.  This is the Application ID URI for the app we are going to be calling, not the one we are login on as.  (See What is the Resource in Azure AD for more information) (optional) Tenant Id : This is used to figure out where the Token Endpoint is.  The new Azure portal calls this a Directory ...
Read more

Querying for items in an Array in CosmosDB

If you have spent any time looking at the documentation for Microsoft CosmosDB / DocumentDB, you will see a lot of examples where the data model has a property named "Tags" that is a list of strings.  But you don't see many times they query on something in that Tag property...  One example I saw a query on Tags[0] = "some value" I don't know how often I will need that, but you know, good to know you can do it. After looking through the SQL syntax reference .  The 2 ways I most likely query the Tags would be to use a join on the Tags property or use the ARRAY_CONTAINS function. Side note; the performance of the two methods are basically identical, leading me to believe the query optimizer generates the same instruction sets for both. So unless you have an array of complex objects, just use ARRAY_CONTAINS. Cool, we know how to query for documents that have our tag on them now... One small problem, when you load a million, or even a hundred thousand do...
Read more

What is the Resource in Azure AD OAuth2

In my post about getting the Azure AD JWT via Postman , we kind of skipped over the part on what the resource is when requesting a token.  Basically, its the Application Id URI of the app you will be calling, not the app you are login in as.  Let's talk through a scenario, using the https://github.com/Azure-Samples/active-directory-dotnet-daemon sample from my post on Service to Service auth. You would create two applications in Azure AD, and they don't even need to be in the same AD (we will cover that at a later date).  The service that is going to consume the API (TodoListDaemon) is the one that would be using the method I described to get a bearer token, and it would have its own Application as we would use its App ID for the Client Id.  Then the service hosting the API (TodoListService) would also have its own Application, but we don't need to generate an Application Key, as we are not logging on as this application.  But in TodoListService we would p...
Read more