Propose a simple trigger prioritization algo for flex event #1112
Conversation
flexible_event_config.md
Outdated
|
|
||
| Given that triggering attribution can affect a source's state without producing a report, we will need a new algorithm for doing trigger prioritization. Here is a sketch of how it could work: | ||
|
|
||
| 1. Always maintain a sorted list of triggers, sorted in order of priority (descending) and trigger time (ascending) |
There was a problem hiding this comment.
Always maintain a sorted list of triggers that are not fully consumed in sent report , sorted in order of priority (descending), then trigger time (ascending).
There was a problem hiding this comment.
What does "fully consumed" mean here?
There was a problem hiding this comment.
Trigger value was fully applied in the creation of a report whose report time is earlier than the trigger time (for "count" operator, the value is 1).
There was a problem hiding this comment.
(Current trigger time, so report time less than current window start).
There was a problem hiding this comment.
In the context of this PR, I don't understand how a trigger could be partially consumed. Not that this PR goes into the details of summary buckets, but in step 2.1 is the trigger's entire value not added to the spec's current value, resulting in N reports depending on the summary buckets?
There was a problem hiding this comment.
@yanzhangnyc I think you're describing an implementation, not an explainer-level algorithm description, particularly with the idea that part of a trigger's value is left over for another report in your third item.
There was a problem hiding this comment.
Why isn't deleting all triggers associated with the spec + window sufficient?
Let's say we consumed 10 of a trigger's value of 12. If we erase this trigger, we have no way of knowing the priority of the remaining value of 2 as it applies to the procedure of generating reports in the next window.
There was a problem hiding this comment.
Why isn't deleting all triggers associated with the spec + window sufficient?
Let's say we consumed 10 of a trigger's value of 12. If we erase this trigger, we have no way of knowing the priority of the remaining value of 2 as it applies to the procedure of generating reports in the next window.
That is an implementation detail not relevant to the explainer-level description here and is covered by "update the source's state based on all of the triggers that were successfully applied."
There was a problem hiding this comment.
Ok, then if nothing else, this shows how hard it is for me to understand the current proposal.
There was a problem hiding this comment.
What does "update the source's state based on all of the triggers that were successfully applied" mean?
flexible_event_config.md
Outdated
| Given that triggering attribution can affect a source's state without producing a report, we will need a new algorithm for doing trigger prioritization. Here is a sketch of how it could work: | ||
|
|
||
| 1. Always maintain a sorted list of triggers, sorted in order of priority (descending) and trigger time (ascending) | ||
| 2. Whenver a spec's window's end time is hit (breaking ties arbitrarily) |
There was a problem hiding this comment.
When a new trigger "matching" this source, we will perform the following process. Not at end of a reporting window
There was a problem hiding this comment.
Why does this matter?
There was a problem hiding this comment.
I intentionally described this per reporting window to minimize the amount of state / accounting. I think the end result should be the same?
There was a problem hiding this comment.
To me it's a lot easier to understand this algorithm when it's described as happening at the end of each report window, not when a trigger is handled.
There was a problem hiding this comment.
@csharrison @giladbarkan-github , Depending on how to undertand "end of trigger window". For example, at end of the trigger window of trigger_data 1, does 1) all triggers processed or 2) only triggers containing trigger_data 1 is processed? If 2), the trigger data with earlier reporting window will have the highest priority. If 1), it might be even more complicated than handling it at trigger time. Let me think about any other complication for this option.
There was a problem hiding this comment.
@yanzhangnyc The PR says "iterate through triggers in order". I take that to mean all triggers associated with this source.
There was a problem hiding this comment.
@yanzhangnyc The PR says "iterate through triggers in order". I take that to mean all triggers associated with this source.
If so, we need to emphasize "at the end window of any trigger data" instead of just simply mention "trigger window end".
There was a problem hiding this comment.
I updated the wording. PTAL
flexible_event_config.md
Outdated
| Given that triggering attribution can affect a source's state without producing a report, we will need a new algorithm for doing trigger prioritization. Here is a sketch of how it could work: | ||
|
|
||
| 1. Always maintain a sorted list of triggers, sorted in order of priority (descending) and trigger time (ascending) | ||
| 2. Whenver a spec's window's end time is hit (breaking ties arbitrarily) |
There was a problem hiding this comment.
To me it's a lot easier to understand this algorithm when it's described as happening at the end of each report window, not when a trigger is handled.
Co-authored-by: Andrew Paseltiner <[email protected]>
Co-authored-by: Andrew Paseltiner <[email protected]>
| Given that triggering attribution can affect a source's state without producing a report, we will need a new algorithm for doing trigger prioritization. Here is a sketch of how it could work: | ||
|
|
||
| 1. For every source, maintain a list of triggers, sorted in order of priority (descending), then trigger time (ascending) | ||
| 2. At the end of any report window (across all of a source's specs, breaking ties arbitrarily): |
There was a problem hiding this comment.
Is there any concern that breaking ties arbitrarily will make it harder to ensure interop conformance?
There was a problem hiding this comment.
Could someone please give an example where breaking a tie one way would differ from another? What choice are we referencing here?
apasel422
added
the
flexible-event
| 2. At the end of any report window (across all of a source's specs, breaking ties arbitrarily): | ||
| 1. Iterate through the source's triggers in order, "applying" them to generate a list of "speculative" reports. Stop whenever privacy limits are hit. | ||
| 2. Send all of the speculative reports that are scheduled to be emitted in the current window | ||
| 3. Update the source's total value per trigger data, and total # of event-level reports based on all of the triggers that were successfully applied in the current window. |
There was a problem hiding this comment.
What's the difference between "successfully" applied and otherwise applied?
| 2. Send all of the speculative reports that are scheduled to be emitted in the current window | ||
| 3. Update the source's total value per trigger data, and total # of event-level reports based on all of the triggers that were successfully applied in the current window. | ||
| 4. Delete all of the speculative reports that are not scheduled to be emitted in this window. | ||
| 5. Delete all of the triggers associated with the current spec and window from the source's trigger list |
There was a problem hiding this comment.
Not sure if this works with the current Android implementation that has "partial" application of a trigger's value, which can extend to the next window. Maybe I'm misunderstanding.
| 1. Iterate through the source's triggers in order, "applying" them to generate a list of "speculative" reports. Stop whenever privacy limits are hit. | ||
| 2. Send all of the speculative reports that are scheduled to be emitted in the current window | ||
| 3. Update the source's total value per trigger data, and total # of event-level reports based on all of the triggers that were successfully applied in the current window. | ||
| 4. Delete all of the speculative reports that are not scheduled to be emitted in this window. |
There was a problem hiding this comment.
How did the reports that are deleted here get created?
Fixes #900