DDoS-ing the Metadata Server.

[xvzf]

Hello,

With high volume traffic, a check (on each interaction with Google APIs!!!) if the application is running on GCP by calling the metadata server is becoming quite a bottleneck response-time wise.

The specific code is located here:

google-auth-library-php/src/Credentials/GCECredentials.php

Line 258 in c438cda

public static function onGce(callable $httpHandler = null)

Environment details

• OS: Linux
• PHP version: 7.x
• Package name and version: irrelevant

[xvzf]

@dwsupplee since you tagged it as investigating, it's not just PHP 7.4 but all 7.x releases

[ericnorris]

At Etsy we've encountered the same issue; during periods of high enough traffic (e.g. backfills) we've seen metadata server outages that result in the following backtrace:

Could not load the default credentials. Browse to https://developers.google.com/accounts/docs/application-default-credentials for more information
DomainException trace:
[Dev_Backtrace]:
#12 getCredentials() /GoogleCloudCredentials.php:70
...

...where the getCredentials is this call:

google-auth-library-php/src/ApplicationDefaultCredentials.php

Line 148 in c05c0a5

public static function getCredentials(

...which ends up calling the onGce method in the original comment.

We use the FetchAuthTokenCache to avoid hammering the metadata server with auth calls, but this onGce call went unnoticed until we experienced outages. Internally we are trying to resolve this by passing in a special $httpHandler to ApplicationDefaultCredentials::getCredentials, but ideally the library would offer a PSR-6 way of caching this, much like how the FetchAuthTokenCache works.

[bshaffer]

If we added a new class, like @ericnorris suggested, we could have something like this:

// GCE cache implementing CacheItemPoolInterface
$cache = new FileCache();
// optional GCE cache config
$cacheConfig = [
    'lifetime' => 1500,
    'prefix' => '',
];
$gceCache = new GCECache($cache, $cacheConfig);

// fetch default credentials with the cache
$credentials = ApplicationDefaultCredentials::getCredentials($scope, null, null, null, null, $gceCache);
// check directly with the cache
$onGce = GCECredentials::getCredentials(null, $gceCache);

We can't use the existing $cache and $cacheConfig parameters passed to ApplicationDefaultCredentials::getCredentials because the cache config / cache interface may be different for the GCE metadata cache.

We could also add CacheItemPoolInterface and $cacheConfig as additional parameters and use GCECache under the hood. This would be messier, but at least it would be consistent:

// FetchAuthTokn cache implementing CacheItemPoolInterface
$authTokenCache = new FileCache();
// optional GCE cache config
$authTokenCacheConfig = [
    'lifetime' => 1500,
    'prefix' => '',
];

// GCE cache implementing CacheItemPoolInterface
$gceCache = new FileCache();
// optional GCE cache config
$gceCacheConfig = [
    'lifetime' => 1500,
    'prefix' => '',
];

// fetch default credentials with the cache
$credentials = ApplicationDefaultCredentials::getCredentials(
    $scope,
    null,
    null,
    $authTokenCache,
    $authTokenCacheConfig,
    $gceCache,
    $gceCacheConfig
);
// check directly with the cache
$onGce = GCECredentials::getCredentials(null, $gceCache, $gceCacheConfig);

Do these solutions sound sufficient, and do you have a preference for either of them?

[ericnorris]

@bshaffer, I actually proposed a different solution internally (props to @etsteeeve for implementing) that seems to work for us. We've instead written a wrapper class called LazyApplicationDefaultCredentials that only calls ApplicationDefaultCredentials::getCredentials when ::fetchAuthToken and ::getLastReceivedToken is called on it. Since this is passed into a FetchAuthTokenCache, this means the ::onGce call is only made once(ish) every ~25 minutes or so.

Perhaps this could be implemented in the library as a DefaultCredentialsFetcher?

[bshaffer]

@ericnorris This is a really great idea! If you or @etsteeeve want to submit a PR please do! Otherwise I can take a look.

It would be nice to have caching for ::onGce independently, so we could implement both of these options.