feat: fastly compute service to perform assignments (FF-3552) #69

leoromanovsky · 2024-11-16T05:04:16Z

Motivation

Eppo is launching a server-based assignment service hosted on the Fastly Compute platform. This is an edge-product that makes

Design details available in the internal doc.

PR goals

Create a service that is able to respond to POST /assignments with a subjectKey and subjectAttributes and return assignments for active allocations.

PR non-goals

This work does not include integration testing or CICD integration beyond compilation. This functionality will be added before a 1.0.0 release in the coming weeks.

Description

Service responds to requests on GET /health and POST /assignments?apiKey - the apiKey format is re-used for legacy reasons although the marketing copy is SDK Key.
The incoming apiKey is converted to a hash and used to look up items in the attached KV store. An external process populates the KV store on Flag changes.
The handler performs assignment across all available flags in series and returns the ones with success evaluation. Any errors in evaluation are logged and the flags are not included in the response.

…l assignments for a given subject against the cached flags (FF-3571)

fastly-edge-assignments/src/handlers/assignments.rs

…mpute all assignments for a given subject against the cached flags (FF-3571)" This reverts commit a159065.

leoromanovsky · 2024-11-21T00:30:22Z

Makefile

+
+# Build only the `fastly-edge-assignments` package for WASM
+.PHONY: fastly-edge-assignments-build
+fastly-edge-assignments-build:


Are these changes helpful? I got tired of putting the various parameters in the right place.

leoromanovsky · 2024-11-21T00:31:07Z

fastly-edge-assignments/fastly.toml

+  [local_server.kv_stores]
+    [[local_server.kv_stores.edge-assignment-kv-store]]
+      key = "ufc-by-sdk-key-token-hash-V--77TScV5Etm78nIMTSOdiroOh1__NsupwUwsetEVM"
+      file = "../sdk-test-data/ufc/flags-v1.json"


this instructs the fastly serve process to populate the local in-memory KV store with one key, a UFC configuration at a precomputed hash. See the unit test in this repo for the associate SDK key to use.

eppo_core/src/ufc/assignment.rs

fastly-edge-assignments/src/handlers/assignments.rs

rasendubi · 2024-11-21T11:59:31Z

fastly-edge-assignments/src/handlers/assignments.rs

+    allocation_key: String,
+    variation_key: String,
+    variation_type: String,
+    variation_value: serde_json::Value,


minor: inside core we have ValueWire that represents this better

fastly-edge-assignments/src/handlers/assignments.rs

rasendubi · 2024-11-21T12:05:19Z

fastly-edge-assignments/src/handlers/assignments.rs

+}
+
+const KV_STORE_NAME: &str = "edge-assignment-kv-store";
+const SDK_KEY_QUERY_PARAM: &str = "apiKey"; // For legacy reasons this is named `apiKey`


minor: given this is a completely new service, I think we can go with sdkKey as there are no legacy clients?

I want to do this as well - @sameerank used our existing javascript client to access this service to setup a connection for our eppo.cloud QA environment so we can start collect telemetry on real-world use. Eppo-exp/js-sdk-common#137

I have a couple of goals to evolve the network architecture of this service:

Serve it behind our existing CDN (with no caching) to be able to share the same domain name and SSL certificate

Since Sameeran developed the new client separately from the old we have a lot of control over the query parameters and should make this change; or even better, pass the auth information in a Authorization header.

rasendubi · 2024-11-21T12:14:12Z

fastly-edge-assignments/src/handlers/assignments.rs

+    base64_url::encode(&hasher.finalize())
+}
+
+pub fn handle_assignments(mut req: Request) -> Result<Response, Error> {


minor: this function has potential to be split into multiple helper functions, which is currently hindered by the fact that error responses are constructed inline. If we have an explicit error type, that would make splitting much easier and the main body more pleasant and straightforward

Ok I see what you mean - ideally we return explicit errors and this handle just is responsible to rendering those into various http codes and messages?

fastly-edge-assignments/src/main.rs

leoromanovsky · 2024-11-23T03:34:35Z

eppo_core/src/precomputed_assignments.rs

+
+impl FlagAssignment {
+    pub fn try_from_assignment(assignment: Assignment) -> Option<Self> {
+        // WARNING! There is a problem here. The event is only populated for splits


@rasendubi @sameerank I believe this is a problem as described here: https://github.com/Eppo-exp/eppo-multiplatform/blob/main/eppo_core/src/ufc/compiled_flag_config.rs#L231

During compilation, an optimization step to reduce the configuration, event is only populated for splits with do_log == True.

A possible solution is to produce a configuration that is "not optimized" for use in this function so that we can access do_log

Another option is to make fields from event optional. There's no reason to include keys and extra logging if doLog is false.

I still think its a problem because we want the precomputed server to return evaluations for all flags and do_log is true or false depending on if the user wants them log, but they still need to be present. The flag will be present in the response which changes the behavior of the SDK / application instead of now the flag will not be present and the default value will be used. Am I understanding that correctly?

leoromanovsky · 2024-11-23T03:42:17Z

eppo_core/src/ufc/models.rs

 pub(crate) struct UniversalFlagConfigWire {
    /// When configuration was last updated.
    pub created_at: Timestamp,
+    pub format: AssignmentFormat,


heads up @sameerank I added a new key to UFC parsing. I see it is being returned now always so I feel it is safe to make it required.

leoromanovsky · 2024-11-23T03:43:11Z

fastly-edge-assignments/src/handlers/assignments.rs

+    base64_url::encode(&hasher.finalize())
+}
+
+pub fn handle_assignments(mut req: Request) -> Result<Response, Error> {


Ok I see what you mean - ideally we return explicit errors and this handle just is responsible to rendering those into various http codes and messages?

leoromanovsky · 2024-11-23T04:36:51Z

@sameerank I'm going to merge this PR even though we still have some feedback from Oleksii and improvements I pointed out in the comments, to have a milestone to match your JS SDK release; the do_log issue is most pressing to figure out.

rasendubi · 2024-11-25T16:56:43Z

eppo_core/src/precomputed_assignments.rs

+            extra_logging: event
+                .base
+                .extra_logging
+                .iter()
+                .map(|(k, v)| (k.clone(), v.clone()))
+                .collect(),


same as:

extra_logging: event.base.extra_logging.clone(),

rasendubi · 2024-11-25T16:57:37Z

eppo_core/src/precomputed_assignments.rs

+}
+
+impl FlagAssignment {
+    pub fn try_from_assignment(assignment: Assignment) -> Option<Self> {


minor: prefer implementing From and TryFrom traits

rasendubi · 2024-11-25T17:04:49Z

fastly-edge-assignments/src/handlers/assignments.rs

+    let subject_assignments = flag_keys
+        .iter()
+        .filter_map(|key| {
+            match evaluator.get_assignment(key, &subject_key, &subject_attributes, None) {
+                Ok(Some(assignment)) => FlagAssignment::try_from_assignment(assignment)
+                    .map(|flag_assignment| (key.clone(), flag_assignment)),
+                Ok(None) => None,
+                Err(e) => {
+                    eprintln!("Failed to evaluate assignment for key {}: {:?}", key, e);
+                    None
+                }
+            }
+        })
+        .collect::<HashMap<_, _>>();
+
+    // Create the response
+    let assignments_response = PrecomputedAssignmentsServiceResponse::from_configuration(
+        configuration,
+        subject_assignments,
+    );


major: computing subject assignments should be part of core. super-weird to have request/response formats in the core but the way to calculate assignment not in the core

rasendubi · 2024-11-25T17:05:56Z

eppo_core/src/precomputed_assignments.rs

+// Request
+#[derive(Debug, Deserialize)]
+pub struct PrecomputedAssignmentsServiceRequestBody {


major: eppo_core shall be unconcerned with network request format

rasendubi · 2024-11-25T17:08:33Z

eppo_core/src/ufc/models.rs

 pub(crate) struct UniversalFlagConfigWire {
    /// When configuration was last updated.
    pub created_at: Timestamp,
+    pub format: AssignmentFormat,


minor: ideally, this field should be optional because there could exist configurations without it (remember that users may be storing and passing configurations out of band now)

leoromanovsky added 13 commits November 15, 2024 21:03

feat: fastly compute service to perform assignments (FF-3552)

a813273

convert from reqwest blocking to non-blocking to conform to wasm target

e985863

remove vscode settings

c8717ee

makefile for local development

5956e65

split into multiple handlers

a3f37c7

its a thing of beauty

48d47aa

coming along nicely

56e5c96

good

3a4b644

local test kv-store!

59cb251

feat: add get_subject_assignments method to eppo_core to compute al…

a159065

…l assignments for a given subject against the cached flags (FF-3571)

simplify unnecessary ownership transfer

6e704ef

e2e evaluation

bccce04

remove space end of filename

c338d31

rasendubi reviewed Nov 18, 2024

View reviewed changes

fastly-edge-assignments/src/handlers/assignments.rs Outdated Show resolved Hide resolved

leoromanovsky added 6 commits November 18, 2024 13:14

Revert "feat: add get_subject_assignments method to eppo_core to co…

ff73c47

…mpute all assignments for a given subject against the cached flags (FF-3571)" This reverts commit a159065.

tidy up error logging

48d2ebc

Merge branch 'main' into lr/ff-3552/edge-function

3b55f6d

edge function dep on 4.1.1

3219a03

generate token hash

d288696

fastly local

be2d5ac

leoromanovsky marked this pull request as ready for review November 21, 2024 00:29

leoromanovsky commented Nov 21, 2024

View reviewed changes

leoromanovsky requested review from greghuels, rasendubi and sameerank November 21, 2024 00:31

conform to desired API

9f99865

rasendubi approved these changes Nov 21, 2024

View reviewed changes

sameerank reviewed Nov 22, 2024

View reviewed changes

fastly-edge-assignments/src/main.rs Show resolved Hide resolved

leoromanovsky added 2 commits November 22, 2024 11:35

add CORS support

78fe3b5

Merge branch 'main' into lr/ff-3552/edge-function

f156caa

leoromanovsky added 15 commits November 22, 2024 16:29

version from cargo

f2f27af

use Datetime

f1bcf3d

String

4ada237

no copy

8101988

use VariationType

50ed0ec

move structs to core

38f71eb

extra logging

266de8a

refactor FlagAssignment to model

c3c90f8

use Str

b9ca60a

build responses in core

26300db

tidy

26c4466

add format

7b4eb91

str

95219d4

format

81d1f55

one more format

9622a21

leoromanovsky commented Nov 23, 2024

View reviewed changes

leoromanovsky assigned rasendubi and sameerank Nov 23, 2024

leoromanovsky merged commit 760d58d into main Nov 23, 2024
25 checks passed

leoromanovsky deleted the lr/ff-3552/edge-function branch November 23, 2024 04:37

rasendubi reviewed Nov 25, 2024

View reviewed changes

leoromanovsky mentioned this pull request Nov 25, 2024

Make UniversalFlagConfigWire.format optional #76

Merged

feat: fastly compute service to perform assignments (FF-3552) #69

feat: fastly compute service to perform assignments (FF-3552) #69

Uh oh!

Conversation

leoromanovsky commented Nov 16, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

PR goals

PR non-goals

Description

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

leoromanovsky commented Nov 23, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

leoromanovsky commented Nov 16, 2024 •

edited

Loading

leoromanovsky commented Nov 23, 2024 •

edited

Loading