Think of it as getting a discount on $50,000 worth of tokens. The way you run your process will pull from that bucket as if you had $50,000 to spend even though you are paying $46,000. You have a year or 3 years to spend them, if you don't spend them in that timeframe you lose the remaining. If you go through them all you will pay the pay-as-you-go price or you can pre-buy another year or 3 year bucket of units. Also how you run your jobs and what tier you run under (Standard or Premium) will determine how fast you burn through the bucket of units and does still matter as the previous answer stated.
https://azure.microsoft.com/en-us/pricing/details/databricks/
Databricks Unit pre-purchase plan
You can get up to 37% savings over pay-as-you-go DBU prices when you
pre-purchase Azure Databricks Units (DBU) as Databricks Commit Units
(DBCU) for either 1 or 3 years. A Databricks Commit Unit (DBCU)
normalizes usage from Azure Databricks workloads and tiers into to a
single purchase. Your DBU usage across those workloads and tiers will
draw down from the Databricks Commit Units (DBCU) until they are
exhausted, or the purchase term expires. The draw down rate will be
equivalent to the price of the DBU, as per the table above.
The purchase tiers and discounts for DBCU purchases are shown below:
1-year pre-purchase plan
DATABRICKS COMMIT UNIT (DBCU) PRICE (WITH DISCOUNT) DISCOUNT
25,000 $23,500 6%
50,000 $46,000 8%
100,000 $89,000 11%
200,000 $172,000 14%
350,000 $287,000 18%
500,000 $400,000 20%
750,000 $578,000 22%
1,000,000 $730,000 27%
1,500,000 $1,050,000 30%
2,000,000 $1,340,000 33%
Also Analytics/Engineering/Light are not options that you choose from. They are defined by how you run your jobs. Executing a job through the notebook interface is defined as an Analytics job where as if you schedule the notebook to run that is considered an engineering job and if you use a coded library submit job you are running under the light tier.
UPDATE - not enough room in comment section to answer OP reply
great thanks for your answer! I think I got my mistake, but please approve once again. So DBCU is about US dollars, so 50k DCBUs may be equal to let say ~100k DBUs, right?
DBUs and DBCUs are exactly the same and are charged the same as far as usage. The only difference is that you get an up front discount of 8% with your example of pre buying 50,000. If you were to run everything exactly the same in two different workspaces and you spent exactly 50,000 DBU Hours in one and 50,000 DBCU Hours in the other, you would owe $50,000 over the course of the year or you would pay $46,000 up front. Neither of these include the actual VM base costs that you would owe to Azure. The DBU structure is Databricks cut of the cost, so you would have to factor that in to your overall cost.
This took me a while to figure out when I started with databricks as well. When they say you are charged $0.55 for the Analytical job that is per DBU hour that is processed not .55 per job. So if I run an Analytical job for 1 hour I would burn .55 * (# of VM's * VM DBU cost per hour). If I ran that same job for only 1/2 an hour I would be charged (.55*.5) * (# of VM's * (VM DBU cost*.5)). It's easier to think of the DBU and DBCU units as 1 unit = $1 and you are burning the dollar value per second of compute not the unit count. The pricing grid that shows $0.55/DBU should be labeled $0.55/DBU-hour in my opinion. Took me a long time, a couple calls and a poc, to figure out.
As to your second question
And scheduling jobs through REST API is more beneficial then scheduling through ADF => Notebook, right?
Again the question is more complicated that it seems like it should be. I initially said yes it is better, I didn't catch the ADF portion of the question. You can run engineering jobs through ADF by making use of the job cluster option to run your notebooks. If you attach your notebooks through ADF to a premade analytics cluster you will pay the analytics cost. Using the API's you could schedule your notebooks in the built in jobscheduler that databricks provides. My understanding is that is charged at the engineer level of a Notebook and light level if a job library.
Another thing to ask for when prebuying if you go that route is to be able to attach the bucket of units to both your dev/test environment and prod environment. We keep them completely separate networks so we have two workspaces. can both pull from the same pool of units. Depends on your azure setup. We went through Databricks sales when we set ours up but Microsoft should be able to do the same.
