[overlord-issues] [JBoss JIRA] (ARTIF-683) Switch to RDBMS, JPA (Hibernate), Hibernate Search, and Hibernate 2nd Level Cache as the default persistence solution

[ https://issues.jboss.org/browse/ARTIF-683?page=com.atlassian.jira.plugin.... ] Brett Meyer updated ARTIF-683: ------------------------------ Description: Artificer currently uses ModeShape + Infinispan + JDBC as its storage. Back when Artificer was a simple S-RAMP impl, JCR made a lot of sense. The S-RAMP spec is essentially a hierarchical artifact repo that maintains the node metadata and relationships between them. However, the "hierarchical" bit is overstated -- it's limited to a primary artifact and its derived artifact (ex -- primary: XSD, derived: type declarations). So, the hierarchy is at most 2 levels and could be represented by a simple relationship or one-to-one foreign key. The only time the hierarchical structure is helpful is when we look up an artifact by its UUID (due to a specific tree structure we use). But otherwise, I think it's a bit of a misnomer. We're now extending well beyond S-RAMP. In addition to an artifact/metadata/info repo, we're trying to position the project as a more general repo for multiple projects and service information. Most importantly, the relationship requirements will expand the most. As such, I'm thinking we'd be better served by alternatives. Note that this is essentially a read-intensive system. Writes do of course occur, but they're almost always *additions*. Nodes are rarely updated once created. Locking and isolation should be used, but can be extremely optimistic. Also note that most artifacts have files with them. That currently uses a local filesystem store through ISPN, but could certainly be NAS. Additional fuel for the fire: many enterprise-level development shops have millions of artifacts, exponentially higher once derivation kicks in. Further, many have multiple relationships defined. Ideas: 1.) Switch to RDBMS, JPA (Hibernate), Hibernate Search, and Hibernate 2nd Level Cache. Although the structure originally looked JCR-specific, it may make a lot more sense as a relational DB. HSearch is a no brainer -- the full-text search capability would be vastly improved, right out of the box. And the RDBMS + in-memory-cache would be perfect for the read-intensive environment and scalability. 2.) Graph databases: Neo4j (w/ or w/o Hibernate OGM), OrientDB, etc.. The concern here is mainly horizontal scaling and, from what I understand, their (lack of adequate) clustering support. But, it's definitely an option. 3.) Distributed but strongly consistent database: RocksDB (a variant of LevelDB), CockroachDB. These are newer, but can (theoretically) scale larger than relational, and because they replicate data it might be more durable or at least recover faster in the event of failure. On the other hand, this may be more difficult for enterprises to adopt 3.) Stick with MS + ISPN, but use Cassandra behind it (instead of JDBC). Arguably, this wouldn't really change things and could potentially end up worse. 4.) Tinkerpop/Blueprints (graph API). Hawkular is using this. However, from what I've heard elsewhere, it's a horrible standard. Solutions that attempt to implement it end up in a state of twisted adaptation, resulting in performance hits. In the end, I'd argue that #1 is the best from enterprise-level, scalability, reliability, and configurability standpoints. was: Artificer currently uses ModeShape + Infinispan + JDBC as its storage. Back when Artificer was a simple S-RAMP impl, JCR made a lot of sense. The S-RAMP spec is essentially a hierarchical artifact repo that maintains the node metadata and relationships between them. However, the "hierarchical" bit is overstated -- it's limited to a primary artifact and its derived artifact (ex -- primary: XSD, derived: type declarations). So, the hierarchy is at most 2 levels and could be represented by a simple relationship or one-to-one foreign key. The only time the hierarchical structure is helpful is when we look up an artifact by its UUID (due to a specific tree structure we use). But otherwise, I think it's a bit of a misnomer. We're now extending well beyond S-RAMP. In addition to an artifact/metadata/info repo, we're trying to position the project as a more general repo for multiple projects and service information. Most importantly, the relationship requirements will expand the most. As such, I'm thinking we'd be better served by alternatives. Note that this is essentially a read-intensive system. Writes do of course occur, but they're almost always *additions*. Nodes are rarely updated once created. Locking and isolation should be used, but can be extremely optimistic. Also note that most artifacts have files with them. That currently uses a local filesystem store through ISPN, but could certainly be NAS. Additional fuel for the fire: many enterprise-level development shops have millions of artifacts, exponentially higher once derivation kicks in. Further, many have multiple relationships defined. Ideas: 1.) Switch to RDBMS + Hibernate ORM + Hibernate Search + Hibernate 2nd Level Caching. Although the structure originally looked JCR-specific, it may make a lot more sense as a relational DB. HSearch is a no brainer -- the full-text search capability would be vastly improved, right out of the box. And the RDBMS + in-memory-cache would be perfect for the read-intensive environment and scalability. 2.) Graph databases: Neo4j (w/ or w/o Hibernate OGM), OrientDB, etc.. The concern here is mainly horizontal scaling and, from what I understand, their (lack of adequate) clustering support. But, it's definitely an option. 3.) Distributed but strongly consistent database: RocksDB (a variant of LevelDB), CockroachDB. These are newer, but can (theoretically) scale larger than relational, and because they replicate data it might be more durable or at least recover faster in the event of failure. On the other hand, this may be more difficult for enterprises to adopt 3.) Stick with MS + ISPN, but use Cassandra behind it (instead of JDBC). Arguably, this wouldn't really change things and could potentially end up worse. 4.) Tinkerpop/Blueprints (graph API). Hawkular is using this. However, from what I've heard elsewhere, it's a horrible standard. Solutions that attempt to implement it end up in a state of twisted adaptation, resulting in performance hits. In the end, I'd argue that #1 is the best from enterprise-level, scalability, reliability, and configurability standpoints. -- This message was sent by Atlassian JIRA (v6.3.15#6346)