I recently ran into a strange issue with the 50,000 row query limit that had me puzzled due to the belief that the 50,000 row query limit applied only per-call.
This realization occurred while writing a LWC where the .js file calls apex methods to gather data. Specifically, in this instance, I was creating a custom table that displays some aggregated data and only loads and displays 5 rows at time. Users can page through the table to load and display 5 more items.
Whenever I am loading a table page, the JS side makes 5 separate calls (one for each row) to the same APEX method to gather all the aggregated data. I actually use a Promise to send off all 5 calls at one time.
Inside of this reused APEX method is a few queries that get about 2500 rows combined MAX. This apex method does not query over 2500 rows per-call in any case.
When JS uses a promise to send off 5 of these calls at once, I have no problem. However, I also allow users to download the entire table into a csv. When they hit this button, JS uses the same APEX method to load the data, but this part requires JS to put more than 5 calls to the same method in the Promise that gathers all the data. If the table is 10 pages, it ends up putting 50 calls to the same APEX method in the same promise. This is when I get the 50,000 query limit reached error.
This has me confused because I thought that a single call to an APEX method can't query over 50,000 rows… After playing around a little, I noticed that the error would occur once I tried to load more than 20 rows. 19 rows worked fine because 19 * 2500 < 50,000. But 21 rows errored because 21* 2500 > 50,000.
Therefore, I believe that the 50,000 query limit isn't calculated per APEX Method call, but rather calculated on the total number of simultaneous calls to the same APEX method.
Best Answer
Aura, and LWC, groups several requests together in order to maximize performance. This is called "boxcar'ing" (or boxcarring) in the documentation. All Apex requests in the same boxcar payload are subject to a single set of governor limits, even though they're allowed to fail/succeed independently. This means you need to make sure you don't place DML before callout operation requests, and you'll want to limit the number of requests you perform simultaneously. This is documented behavior and has existed since the beginning of Lightning.
UPDATE: As of Winter 20 release, each request has their own set of Limits applied, even if they are boxcar'ed together Winter 20: Hitting Apex Limits in Server-Side Actions Is More Predictable