Introduce Least Request LB active request bias config #11252

gkleiman · 2020-05-19T00:36:36Z

Commit Message:
Introduce Least Request LB active request bias config

Additional Description:
This patch adds a runtime configurable active request bias to the Least Request Load Balancer. This bias only takes effect when all host weights are not equal. The larger it is, the more aggressively active requests will lower the effective weight of a host.

The least request lb behaves like the round robin lb when the bias is set to 0.0 and all weights are not equal. This is very useful for squeeze/redline tests and fixes the issues described on #10602.

Risk Level: S
Testing: Unit test. Validated on Lyft infra that this approach fixes the issues described in #10602.

…eighted hosts case Signed-off-by: Gastón Kleiman <[email protected]>

mattklein123

LGTM with small comments for a test. Thank you!

/wait

mattklein123 · 2020-05-19T03:07:15Z

source/common/runtime/runtime_features.cc

@@ -81,6 +81,7 @@ constexpr const char* runtime_features[] = {
 constexpr const char* disabled_runtime_features[] = {
    // Sentinel and test flag.
    "envoy.reloadable_features.test_feature_false",
+    "envoy.reloadable_features.alternative_least_request_weights",


If we end up doing a real PR for this, it should be a real config option vs. one of these. I would just use a RuntimeFeatureFlag inside the LB config proto.

mattklein123 · 2020-05-19T03:07:44Z

source/common/upstream/load_balancer_impl.h

+    alternativeWeights_ = Runtime::runtimeFeatureEnabled(
+        "envoy.reloadable_features.alternative_least_request_weights");


For perf reasons I would not re-snap this on every choice. I would just do it in refresh below. I think this is fine.

mattklein123 · 2020-05-19T03:08:02Z

source/common/upstream/load_balancer_impl.h

  const uint32_t choice_count_;
+  bool alternativeWeights_;


alternative_weights_

tonya11en

I recognize I have the unpopular opinion here, but this change seems like a hack to workaround for a lack of operational agility. I think a better approach would be to just add some coefficient (k) that makes the weights look something like:

new_weight = (configured_weight / outstanding_requests^k)

If k=1, things behave as they do today and normalize the weights by dividing it by outstanding requests.
If k=0, you get the behavior in this PR and the outstanding requests don't affect the configured weights.

You also have the freedom to configure anything in between and even make the outstanding requests have a more aggressive impact on the weights if you so choose. This would be trivial to implement, can be runtime configurable, and seems less hacky. WDYT?

tonya11en · 2020-05-19T20:11:31Z

source/common/upstream/load_balancer_impl.h

  const uint32_t choice_count_;
+  bool alternativeWeights_;


Also explain what this means exactly in a comment. It's non-obvious from the name.

mattklein123 · 2020-05-19T20:26:02Z

I think a better approach would be to just add some coefficient (k) that makes the weights look something like:

+1 I like this idea. Let's do that!

Signed-off-by: Gastón Kleiman <[email protected]>

gkleiman · 2020-05-20T01:24:18Z

I think a better approach would be to just add some coefficient (k) that makes the weights look something like:

+1 I like this idea. Let's do that!

Good idea! I implemented this with a global runtime. If things work fine in the experiment, I'll make it a real config option inside LeastRequestLbConfig.

mattklein123

Looks good to test. If the patch looks good I like this idea for a general config flag. Great idea @tonya11en!

/wait

mattklein123 · 2020-05-20T15:44:39Z

source/common/upstream/load_balancer_impl.h

+    ENVOY_LOG(debug, "cluster={} address={} active_requests_exponent={} weight={}",
+              host.cluster().name(), host.address()->asString(), active_requests_exponent_, weight);


definitely trace. I would also probably remove this in the final patch.

mattklein123 · 2020-05-20T15:45:00Z

source/common/upstream/load_balancer_impl.h

+    const double weight = static_cast<double>(host.weight()) /
+                          std::pow(host.stats().rq_active_.value() + 1, active_requests_exponent_);


For perf reasons I would probably check if 1.0 and then branch and avoid the pow call.

Signed-off-by: Gastón Kleiman <[email protected]>

mattklein123

/wait

mattklein123 · 2020-05-20T22:35:56Z

source/common/upstream/load_balancer_impl.h

 #include "common/upstream/edf_scheduler.h"

 namespace Envoy {
 namespace Upstream {

+static const std::string RuntimeLeastRequestsActiveRequestsExponent =


absl::string_view to avoid static fiasco.

Signed-off-by: Gastón Kleiman <[email protected]>

stale · 2020-05-29T03:41:42Z

This pull request has been automatically marked as stale because it has not had activity in the last 7 days. It will be closed in 7 days if no further activity occurs. Please feel free to give a status update now, ping for review, or re-open when it's ready. Thank you for your contributions!

tonya11en · 2020-06-01T19:18:45Z

test/common/upstream/load_balancer_impl_test.cc

+  EXPECT_CALL(random_, random()).WillRepeatedly(Return(0));
+
+  // We should see 2:1 ratio for hosts[1] to hosts[0], regardless of the active request count.
+  hostSet().healthy_hosts_[1]->stats().rq_active_.set(1);


I'm surprised this works. Are the stats affecting the outstanding requests perceived by the LB when the exponent is non-zero?

Yes, the active rq stat is used to calculate the effective weights when the exponent is > 0: https://github.com/envoyproxy/envoy/pull/11252/files/0a2b1fcdad5daeef1136ef7c2208f356d3fcf05f#diff-e746c5f817e7b535220c275b9dea666bR494

stale · 2020-06-09T01:12:21Z

This pull request has been automatically marked as stale because it has not had activity in the last 7 days. It will be closed in 7 days if no further activity occurs. Please feel free to give a status update now, ping for review, or re-open when it's ready. Thank you for your contributions!

tonya11en · 2020-06-09T01:37:30Z

This was just verified to have the desired effect in our infra, so I'm resetting stalebot.

jmarantz · 2020-07-02T17:52:18Z

/wait

Use a plain double instead of a runtime double to store the per-cluster active request bias. Note: The goal of this commit is to evaluate the memory overhead of this approach. A commit with te Least Requests LB changes might follow if we deem the memory overhead of this approach acceptable. Signed-off-by: Gastón Kleiman <[email protected]>

jmarantz · 2020-07-06T23:35:29Z

Per above I'm OK merging what you have temporarily. But I think what we want here is a RuntimeDouble for the process, not one for each cluster. I'm not 100% sure of the best data structure to hold that RuntimeDouble.

Signed-off-by: Gastón Kleiman <[email protected]>

gkleiman · 2020-07-07T21:01:30Z

Per above I'm OK merging what you have temporarily. But I think what we want here is a RuntimeDouble for the process, not one for each cluster. I'm not 100% sure of the best data structure to hold that RuntimeDouble.

@jmarantz 👍 I restored the functional changes and merged master — PerWorkerStatsAndBalancing/IPv4 failed the CI coverage run, but I did over 1000 local runs and they all succeeded (bazel test test/integration:stats_integration_test --test_filter="*PerWorkerStatsAndBalancing*" --runs_per_test=1000).

Signed-off-by: Gastón Kleiman <[email protected]>

gkleiman · 2020-07-08T22:37:16Z

/retest

Looks like Azure Pipelines is degraded and the Linux coverage job failed with the following error:

We stopped hearing from agent i-04515e37cd973b99f. Verify the agent machine is running and has a healthy network connection. Anything that terminates an agent process, starves it for CPU, or blocks its network access can cause this error. For more information, see: https://go.microsoft.com/fwlink/?linkid=846610

repokitteh-read-only · 2020-07-08T22:37:19Z

🐴 hold your horses - no failures detected, yet.

🐱

Caused by: a #11252 (comment) was created by @gkleiman.

see: more, trace.

jmarantz · 2020-07-09T00:00:07Z

/azp run

azure-pipelines · 2020-07-09T00:00:31Z

Azure Pipelines successfully started running 1 pipeline(s).

jmarantz · 2020-07-09T00:02:04Z

Summarizing DM discussion on Slack, we are going to move the RuntimeDouble out of the cluster configuration and into bootstrap. The problem with my earlier suggestion to merge this is that it creates an API at the cluster level that will be hard to remove. @gkleiman indicated he didn't think it would be too bad to drive the config decision from the bootstrap proto.

gkleiman · 2020-07-10T23:44:11Z

I've changed the place/way where the active request bias is stored a few times already since I began working on this, so I would love to reach consensus between Lyft & external maintainers before changing the implementation again.

Here goes a summary of Lyft's requirements, findings about the extra per-cluster memory footprint and possible approaches. I hope this makes the conversation a bit easier — it turned out to be a bit long for a PR comment, so I would be also happy to move this to a Google Doc so people can comment on it.

Lyft Requirements

In order to improve Envoy’s load balancing behavior during squeeze tests, we would like to make it possible for the least request load balancer to ignore the active request count in the case that all host weights are not equal.

We would do configure this change globally for all Envoy instances in our service mesh. Being able to override the active request bias for a particular cluster would be nice to have but is currently not required.

Findings

The stats memory integration tests report that with the current approach (Initial import #1) the per-cluster memory usage increases by 256 bytes. That would be 2.56 MB for an Envoy instance with 10k clusters.
The memory tests don’t create any LBs, so the extra memory usage is attributable to the new field in LeastRequestLbConfig and not to the changes in the LB implementation. Note that even though RuntimeDouble contains a runtime key, the memory tests create any instances of the new field. So the actual memory usage for users of this feature that specify runtime keys would be higher than reported by the test.
When adding three fields to LeastRequestLbConfig the memory tests still report 256 bytes of extra per-cluster memory usage. This hints that the extra memory usage might be related to alignment/padding or other optimizations.
As far as I could tell from looking at the code and using a debugger no instances of LeastRequestLbConfig should be created by the tests — it is still unclear why more memory is used if we’re defining a field but not creating any instances of the message containing it.

Feedback

@tonya11en and @mattklein123 initially proposed the current approach — per-cluster RuntimeDouble in `LeastRequestLbConfig (#1)

@jmarantz suggested making the active request bias a global setting instead of a per-cluster setting (#4)

@alyssawilk suggested using scalars for changing per-cluster values via CDS rather than using runtime overrides. (#2)

Approaches

1. Per-Cluster - RuntimeDouble field in CDS LeastRequestLbConfig (Current implementation)

Pro: Easy to implement. Most flexible approach.
Con: Largest memory usage of all approaches (need to store a runtime key per load balancer). Dependency on a v3 xDS server.

2. Per-Cluster - Scalar field in CDS LeastRequestLbConfig

Pro: Smaller memory footprint than using RuntimeDouble , since there is no need to store runtime keys. However the memory tests still report 256 bytes of extra memory. Very easy to implement.
Con: Dependency on a v3 xDS server. Little flexibility, not using runtime means that any changes have to be made via an xDS server.

3. Global Runtime - Hard-coded runtime key (Original implementation)

We could define a global runtime key that can be used to globally configure the active request bias used by all least request LBs, e.g., upstream.least_request_load_balancer.active_request_bias.

Pro: No additional memory footprint. Very easy to implement. Easy to adopt, no dependency on a v3 xDS server.
Con: Little flexibility, using a hard-coded runtime key makes it harder to use different values for different Envoy instances that share the same runtime layers.

4. Global Runtime - Move the RuntimeDouble out of the cluster configuration and into bootstrap

We could move the RuntimeDouble to the bootstrap config message, e.g., we could move it to Bootstrap.ClusterManager.LbConfig.LeastRequestLbConfig.active_request_bias.

To avoid reducing the additional memory usage we could add a Runtime::Double instance owned by the cluster manager and figure out how to share it among load balancers on different threads.

Pro: Small memory footprint. Easy to adopt, no dependency on a v3 xDS server.
Con: Intermediate flexibility. No way of configuring per-cluster values, but it’s possible to use different runtime keys for different Envoy instances. Implementation is a bit more complex than the other approaches.

jmarantz · 2020-07-11T02:00:54Z

Thanks for the detailed note.

It's clear we'll have to spend some memory to get the per-cluster configurability for this feature. I'm fine if that's the consensus. The amount of memory we are spending was surprising, which I believe is due to Envoy's usage of protobufs as retained data structures. I'm not sure how big a task it would be to move away from protobufs as a retained data structure and use them instead only for transport. But that could help get that memory back and perhaps then some. Obviously that would be outside the scope of the PR, obviously. Perhaps someone could own the task of evaluating that optimization as a follow-up.

In the scheme of things, 2.5M memory is obviously not a lot for a proxy meant to handle 10k clusters, so I'm OK merging this in in its current form. I had pushed back earlier, because I thought maybe the per-cluster config was not strictly needed and would be hard to remove once it was added. Sounds like it would be useful at least.

mattklein123 · 2020-07-13T17:59:46Z

I think we should move this PR forward as-is with the per-cluster increase. If you want to leave some TODOs we can do that but I think the proto issues are out of scope for this PR?

jmarantz · 2020-07-13T18:01:14Z

agreed; happy to see it merged and iterate from there.

This reverts commit a6a285d. Signed-off-by: Gastón Kleiman <[email protected]>

Signed-off-by: Gastón Kleiman <[email protected]>

gkleiman · 2020-07-14T21:01:49Z

@mattklein123 IpVersionsClientType/VhdsIntegrationTest.VhdsOnDemandUpdateHttpConnectionCloses/0 failed the tsan run due to a data race that doesn't seem to be related to this PR.

alyssawilk · 2020-07-14T21:21:42Z

FYI master merge should fix that ( #12067)

…

On Tue, Jul 14, 2020 at 5:02 PM Gastón Kleiman ***@***.***> wrote: @mattklein123 <https://github.com/mattklein123> IpVersionsClientType/VhdsIntegrationTest.VhdsOnDemandUpdateHttpConnectionCloses/0 failed the tsan run due to a data race that doesn't seem to be related to this PR. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#11252 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AELALPO6LGGHSY6522Q7KC3R3TBUZANCNFSM4NERGDPA> .

Signed-off-by: Gastón Kleiman <[email protected]>

* Add support for making Least Requests LB behave like Round Robin in weighted hosts case Signed-off-by: Gastón Kleiman <[email protected]> * Address feedback Signed-off-by: Gastón Kleiman <[email protected]> * Perf/logging improvements Signed-off-by: Gastón Kleiman <[email protected]> * Address feedback and cleanup BUILD file Signed-off-by: Gastón Kleiman <[email protected]> * Make active requests exponent configurable via CDS/runtime Signed-off-by: Gastón Kleiman <[email protected]> * Address feedback Signed-off-by: Gastón Kleiman <[email protected]> * Validate log message Signed-off-by: Gastón Kleiman <[email protected]> * Update cluster memory test golden values Signed-off-by: Gastón Kleiman <[email protected]> * Fix method name Signed-off-by: Gastón Kleiman <[email protected]> * Explicitly initialize active_request_bias_ Signed-off-by: Gastón Kleiman <[email protected]> * Try to make clang-tidy happy Signed-off-by: Gastón Kleiman <[email protected]> * Use unique_ptr instead of optional Signed-off-by: Gastón Kleiman <[email protected]> * Update stats integration test Signed-off-by: Gastón Kleiman <[email protected]> * Check whether memory footprint is reduced without LB changes Signed-off-by: Gastón Kleiman <[email protected]> * Use plain double for active request bias Use a plain double instead of a runtime double to store the per-cluster active request bias. Note: The goal of this commit is to evaluate the memory overhead of this approach. A commit with te Least Requests LB changes might follow if we deem the memory overhead of this approach acceptable. Signed-off-by: Gastón Kleiman <[email protected]> * Revert back to approved implementation using RuntimeDouble Signed-off-by: Gastón Kleiman <[email protected]> * Add extra fields to CDS cluster proto to check memory usage Signed-off-by: Gastón Kleiman <[email protected]> * Revert "Add extra fields to CDS cluster proto to check memory usage" This reverts commit a6a285d. Signed-off-by: Gastón Kleiman <[email protected]> * Add changelog entry Signed-off-by: Gastón Kleiman <[email protected]> Signed-off-by: Kevin Baichoo <[email protected]>

* Add support for making Least Requests LB behave like Round Robin in weighted hosts case Signed-off-by: Gastón Kleiman <[email protected]> * Address feedback Signed-off-by: Gastón Kleiman <[email protected]> * Perf/logging improvements Signed-off-by: Gastón Kleiman <[email protected]> * Address feedback and cleanup BUILD file Signed-off-by: Gastón Kleiman <[email protected]> * Make active requests exponent configurable via CDS/runtime Signed-off-by: Gastón Kleiman <[email protected]> * Address feedback Signed-off-by: Gastón Kleiman <[email protected]> * Validate log message Signed-off-by: Gastón Kleiman <[email protected]> * Update cluster memory test golden values Signed-off-by: Gastón Kleiman <[email protected]> * Fix method name Signed-off-by: Gastón Kleiman <[email protected]> * Explicitly initialize active_request_bias_ Signed-off-by: Gastón Kleiman <[email protected]> * Try to make clang-tidy happy Signed-off-by: Gastón Kleiman <[email protected]> * Use unique_ptr instead of optional Signed-off-by: Gastón Kleiman <[email protected]> * Update stats integration test Signed-off-by: Gastón Kleiman <[email protected]> * Check whether memory footprint is reduced without LB changes Signed-off-by: Gastón Kleiman <[email protected]> * Use plain double for active request bias Use a plain double instead of a runtime double to store the per-cluster active request bias. Note: The goal of this commit is to evaluate the memory overhead of this approach. A commit with te Least Requests LB changes might follow if we deem the memory overhead of this approach acceptable. Signed-off-by: Gastón Kleiman <[email protected]> * Revert back to approved implementation using RuntimeDouble Signed-off-by: Gastón Kleiman <[email protected]> * Add extra fields to CDS cluster proto to check memory usage Signed-off-by: Gastón Kleiman <[email protected]> * Revert "Add extra fields to CDS cluster proto to check memory usage" This reverts commit a6a285d. Signed-off-by: Gastón Kleiman <[email protected]> * Add changelog entry Signed-off-by: Gastón Kleiman <[email protected]> Signed-off-by: scheler <[email protected]>

Add support for making Least Requests LB behave like Round Robin in w…

a1b52fd

…eighted hosts case Signed-off-by: Gastón Kleiman <[email protected]>

mattklein123 assigned mattklein123 and tonya11en May 19, 2020

mattklein123 requested changes May 19, 2020

View reviewed changes

repokitteh-read-only bot added the waiting label May 19, 2020

wgallagher previously approved these changes May 19, 2020

View reviewed changes

tonya11en suggested changes May 19, 2020

View reviewed changes

gkleiman mentioned this pull request May 19, 2020

Add runtime flag to make Weighted Least Request LB behave like Weighted Round Robin LB in weighted mode #10602

Closed

Address feedback

762e61b

Signed-off-by: Gastón Kleiman <[email protected]>

gkleiman dismissed wgallagher’s stale review via 762e61b May 20, 2020 01:20

repokitteh-read-only bot removed the waiting label May 20, 2020

mattklein123 requested changes May 20, 2020

View reviewed changes

repokitteh-read-only bot added the waiting label May 20, 2020

Perf/logging improvements

4cf5f19

Signed-off-by: Gastón Kleiman <[email protected]>

repokitteh-read-only bot removed the waiting label May 20, 2020

mattklein123 requested changes May 20, 2020

View reviewed changes

repokitteh-read-only bot added the waiting label May 20, 2020

Address feedback and cleanup BUILD file

0a2b1fc

Signed-off-by: Gastón Kleiman <[email protected]>

repokitteh-read-only bot removed the waiting label May 20, 2020

mattklein123 added the waiting label May 21, 2020

stale bot added the stale stalebot believes this issue/PR has not been touched recently label May 29, 2020

tonya11en reviewed Jun 1, 2020

View reviewed changes

stale bot removed the stale stalebot believes this issue/PR has not been touched recently label Jun 1, 2020

stale bot added the stale stalebot believes this issue/PR has not been touched recently label Jun 9, 2020

stale bot removed the stale stalebot believes this issue/PR has not been touched recently label Jun 9, 2020

repokitteh-read-only bot added the waiting label Jul 2, 2020

repokitteh-read-only bot removed the waiting label Jul 6, 2020

gkleiman added 2 commits July 6, 2020 18:06

Revert back to approved implementation using RuntimeDouble

a4e7b39

Signed-off-by: Gastón Kleiman <[email protected]>

Merge remote-tracking branch 'origin/master' into alternative-lr-weights

1010883

Signed-off-by: Gastón Kleiman <[email protected]>

Add extra fields to CDS cluster proto to check memory usage

a6a285d

Signed-off-by: Gastón Kleiman <[email protected]>

gkleiman added 2 commits July 13, 2020 11:28

Revert "Add extra fields to CDS cluster proto to check memory usage"

2676928

This reverts commit a6a285d. Signed-off-by: Gastón Kleiman <[email protected]>

Merge remote-tracking branch 'origin/master' into alternative-lr-weights

100c3db

Signed-off-by: Gastón Kleiman <[email protected]>

gkleiman added 2 commits July 14, 2020 15:37

Merge remote-tracking branch 'origin/master' into alternative-lr-weights

cf06a76

Signed-off-by: Gastón Kleiman <[email protected]>

Add changelog entry

eb1b857

Signed-off-by: Gastón Kleiman <[email protected]>

jmarantz approved these changes Jul 15, 2020

View reviewed changes

jmarantz merged commit 9f7d448 into envoyproxy:master Jul 15, 2020

gkleiman deleted the alternative-lr-weights branch August 26, 2020 19:09

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Introduce Least Request LB active request bias config #11252

Introduce Least Request LB active request bias config #11252

gkleiman commented May 19, 2020 •

edited

Loading

mattklein123 left a comment

mattklein123 May 19, 2020

mattklein123 May 19, 2020

mattklein123 May 19, 2020

tonya11en left a comment

tonya11en May 19, 2020

mattklein123 commented May 19, 2020

gkleiman commented May 20, 2020

mattklein123 left a comment

mattklein123 May 20, 2020

mattklein123 May 20, 2020

mattklein123 left a comment

mattklein123 May 20, 2020

stale bot commented May 29, 2020

tonya11en Jun 1, 2020

gkleiman Jun 1, 2020

stale bot commented Jun 9, 2020

tonya11en commented Jun 9, 2020

jmarantz commented Jul 2, 2020

jmarantz commented Jul 6, 2020

gkleiman commented Jul 7, 2020

gkleiman commented Jul 8, 2020

repokitteh-read-only bot commented Jul 8, 2020

jmarantz commented Jul 9, 2020

azure-pipelines bot commented Jul 9, 2020

jmarantz commented Jul 9, 2020

gkleiman commented Jul 10, 2020 •

edited

Loading

jmarantz commented Jul 11, 2020

mattklein123 commented Jul 13, 2020

jmarantz commented Jul 13, 2020

gkleiman commented Jul 14, 2020

alyssawilk commented Jul 14, 2020 via email

		alternativeWeights_ = Runtime::runtimeFeatureEnabled(
		"envoy.reloadable_features.alternative_least_request_weights");

		ENVOY_LOG(debug, "cluster={} address={} active_requests_exponent={} weight={}",
		host.cluster().name(), host.address()->asString(), active_requests_exponent_, weight);

		const double weight = static_cast<double>(host.weight()) /
		std::pow(host.stats().rq_active_.value() + 1, active_requests_exponent_);

Introduce Least Request LB active request bias config #11252

Introduce Least Request LB active request bias config #11252

Conversation

gkleiman commented May 19, 2020 • edited Loading

mattklein123 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tonya11en left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mattklein123 commented May 19, 2020

gkleiman commented May 20, 2020

mattklein123 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mattklein123 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

stale bot commented May 29, 2020

Choose a reason for hiding this comment

Choose a reason for hiding this comment

stale bot commented Jun 9, 2020

tonya11en commented Jun 9, 2020

jmarantz commented Jul 2, 2020

jmarantz commented Jul 6, 2020

gkleiman commented Jul 7, 2020

gkleiman commented Jul 8, 2020

repokitteh-read-only bot commented Jul 8, 2020

jmarantz commented Jul 9, 2020

azure-pipelines bot commented Jul 9, 2020

jmarantz commented Jul 9, 2020

gkleiman commented Jul 10, 2020 • edited Loading

Lyft Requirements

Findings

Feedback

Approaches

1. Per-Cluster - RuntimeDouble field in CDS LeastRequestLbConfig (Current implementation)

2. Per-Cluster - Scalar field in CDS LeastRequestLbConfig

3. Global Runtime - Hard-coded runtime key (Original implementation)

4. Global Runtime - Move the RuntimeDouble out of the cluster configuration and into bootstrap

jmarantz commented Jul 11, 2020

mattklein123 commented Jul 13, 2020

jmarantz commented Jul 13, 2020

gkleiman commented Jul 14, 2020

alyssawilk commented Jul 14, 2020 via email

gkleiman commented May 19, 2020 •

edited

Loading

gkleiman commented Jul 10, 2020 •

edited

Loading