Use tiered cache in place of Caffeine cache

[pditommaso]

Tiered cache provides twofolds benefits: 1) the cache content is shared across all replicas, and 2) the cache content is should not be reloaded in across restarts.

The goal of this issue is to replace the local Caffeine cache with a tiered cache for the following services:

• AwsEcrService
• RegistryLookupService
• RegistryAuthService

Supporting those service require however extending the current tiered cache API adding the ability to use Java object a cache keys other than plain strings as it is now, see here

abstract class AbstractTieredCache implements TieredCache<String,V>

This could be achieved introducing an interface like the following one

interface TieredCacheKey {

  String stableHash()

}

The concrete key class should take care of proving an implementation for stableHash, likely using sipHash algorithm.

[munishchouhan]

@pditommaso does tieredcache also addresses #747?
because in AwsEcrService we are using cache.synchronous() to avoid concurreny issue

wave/src/main/groovy/io/seqera/wave/service/aws/AwsEcrService.groovy

Lines 142 to 146 in d5c086d

	// get the token from the cache, if missing the it's automatically
	// fetch using the AWS ECR client
	// FIXME https://github.com/seqeralabs/wave/issues/747
	return cache.synchronous().get(new AwsCreds(accessKey,secretKey,region,isPublic))
	}

[pditommaso]

No, it uses always Ceffeine behind the scene. We need Java 24 to revert that change

[munishchouhan]

Since now we need to use string as key, there is an issue with loader

wave/src/main/groovy/io/seqera/wave/service/aws/AwsEcrService.groovy

Lines 88 to 93 in 1e13c9f

	String load(AwsCreds creds) throws Exception {
	return creds.ecrPublic
	? getLoginToken1(creds.accessKey, creds.secretKey, creds.region)
	: getLoginToken0(creds.accessKey, creds.secretKey, creds.region)
	}
	}

here loader is using key (AwsCreds) to compute the value, changing creds to string will not let us compute it anymore.

[munishchouhan]

we can store awscreds as it is and generate token, when fetchs from cache

[pditommaso]

AwsCreds should implement TieredCacheKey as I was suggesting, then should be the same

[munishchouhan]

yes, that i did here:

wave/src/main/groovy/io/seqera/wave/service/aws/AwsEcrService.groovy

Lines 69 to 77 in 1e13c9f

	String stableHash() {
	final h = Hashing.sipHash24().newHasher()
	for( Object it : [accessKey, secretKey, region, ecrPublic] ) {
	if( it!=null )
	h.putUnencodedChars(it.toString())
	h.putUnencodedChars('/')
	}
	return h.hash()
	}

but loader requires AwsCreds and not a string because it needs compute token from it

[pditommaso]

This needs to be changed to take TieredCacheKey as key instead of string

wave/src/main/groovy/io/seqera/wave/store/cache/AbstractTieredCache.groovy

Line 48 in 67f91c8

abstract class AbstractTieredCache<V extends MoshiExchange> implements TieredCache<String,V> {

ei.

abstract class AbstractTieredCache<K extends TieredCacheKey, V extends MoshiExchange> implements TieredCache<K,V> {

  .. 
}

make sense ?

[munishchouhan]

ok understood, thanks

[munishchouhan]

Another issue is Caffeine cache loader is not a Funtion<>, it doesn't have apply Method

[pditommaso]

It should not impact, the trick is to replace in the internal implementation the string key with TieredCacheKey.stableHash