3 ms·

"At that point you find yourself asking why didn't we just use a NoSQL store" Because you still have the full ability to write general SQL within the shard. Fo

"At that point you find yourself asking why didn't we just use a NoSQL store" Because you still have the full ability to write general SQL within the shard. For many types of application this is useful.

threeseed 13y ago
Sure. But only if ALL the data within that query exists on the shard. It is rare that this would happen if you have anything resembling a normalised schema. In which case you would still be doing a lot of joins in your application layer. IMHO Sharded MySQL very much belongs in the NoSQL camp.
- twic 13y ago
  I don't know that this is true. Imagine you're LeanKit, or Fog Creek, and you run a kanban board as a service. Or a bug tracker, CMS, whatever. You have many customers, each of whom has no more than thousands of users and millions of items. There are many relationships between objects belonging to a given customer, but precisely zero relationships between objects belonging to different customers. Shard using the customer identity as a key, and you have nicely spread-out data and the ability to do any query the application might need to, while still having a normalised schema. There are plenty of other application whose schemas have this property, or almost have it. In my company, we make financial applications, and a lot of the data has very similar siloed ownership structure. The one thing you can't do is reporting queries across your customers. That doesn't seem like a killer, though - it's normal to farm that stuff out to an offline reporting database even in single-server environments.
h1karu 13y ago
That's a fair point. If you need highly relational queries within a given shard then using a relational database there can make sense, but for many apps the kinds of queries that are used are not highly relational in nature so using a NoSQL store could have reduced the need for all of that app-specific sharding engineering while still allowing for all the queries that were needed by the app. I personally find that SQL relational databases are not flexible enough for my application so I go with a polyglot persistence architecture that ties together a nosql key-value store with a graph database technology. This way I can use the cypher query language to access data relationally across all my shards while scaling writes and Big-data horizontally. To each his own.
- balfirevic 13y ago
  "This way I can use the cypher query language to access data relationally across all my shards while scaling writes and Big-data horizontally" That sounds pretty powerful. Do you mind sharing more information about the data storage/querying stack you use?
  - h1karu 13y ago
    NDA and all of that, but if you pick up this book: http://pragprog.com/book/rwdata/seven-databases-in-seven-weeks http://pragprog.com/book/rwdata/seven-databases-in-seven-wee... and grep for the phrase "polyglot persistence" you'll find some great information to get you started. a video promo for the book: https://www.youtube.com/watch?v=bSAc56YCOaE https://www.youtube.com/watch?v=bSAc56YCOaE
    - twic 13y ago
      Can you talk about what sort of queries you do in Neo4j? Graphs are perhaps the only area where NoSQL may beat relational databases for expressiveness, rather than alleged scalability or ease of use. I've read the documentation for Neo4j and Cypher, but never used them in anger, so i'd be really interested to know more about how they are actually used. In particular, i would love to be able to make a comparison between a graph database and a relational database queried using recursive common table expressions, in the context of a real-world problem.
      - h1karu 13y ago
        The approach I like is to use a key/value document store as the system of record (SOR). If any data conflicts arise between Redis, the document store, and the graph database, then the SOR wins. A good SOR should contain all of the data necessary to rebuild any other data source in its domain. Graph databases can scale write throughput fine when those writes are being fed to it from a single source so it's best to have a service who's sole purpose is to keep the graph database in sync with the SOR. The graph should only store the data you actually need in order to get the queries that you want. The specific queries in my domain are exactly the kind of thing I'm not supposed to talk about, but I will say you can do things like: Give me 50 users who who live near me (lat/long bounding box), who I'm not already following, who I have not already sent a message to, who have at least two friends who also live near me, and who have matching tags, order by number of followers. there are a bunch of great examples in this awesome book: http://www.amazon.com/Graph-Databases-Ian-Robinson/dp/1449356265 http://www.amazon.com/Graph-Databases-Ian-Robinson/dp/144935...

threeseed 13y ago

Sure. But only if ALL the data within that query exists on the shard. It is rare that this would happen if you have anything resembling a normalised schema. In which case you would still be doing a lot of joins in your application layer. IMHO Sharded MySQL very much belongs in the NoSQL camp.

twic 13y ago
I don't know that this is true. Imagine you're LeanKit, or Fog Creek, and you run a kanban board as a service. Or a bug tracker, CMS, whatever. You have many customers, each of whom has no more than thousands of users and millions of items. There are many relationships between objects belonging to a given customer, but precisely zero relationships between objects belonging to different customers. Shard using the customer identity as a key, and you have nicely spread-out data and the ability to do any query the application might need to, while still having a normalised schema. There are plenty of other application whose schemas have this property, or almost have it. In my company, we make financial applications, and a lot of the data has very similar siloed ownership structure. The one thing you can't do is reporting queries across your customers. That doesn't seem like a killer, though - it's normal to farm that stuff out to an offline reporting database even in single-server environments.

h1karu 13y ago

That's a fair point. If you need highly relational queries within a given shard then using a relational database there can make sense, but for many apps the kinds of queries that are used are not highly relational in nature so using a NoSQL store could have reduced the need for all of that app-specific sharding engineering while still allowing for all the queries that were needed by the app. I personally find that SQL relational databases are not flexible enough for my application so I go with a polyglot persistence architecture that ties together a nosql key-value store with a graph database technology. This way I can use the cypher query language to access data relationally across all my shards while scaling writes and Big-data horizontally. To each his own.

balfirevic 13y ago
"This way I can use the cypher query language to access data relationally across all my shards while scaling writes and Big-data horizontally" That sounds pretty powerful. Do you mind sharing more information about the data storage/querying stack you use?
- h1karu 13y ago
  NDA and all of that, but if you pick up this book: http://pragprog.com/book/rwdata/seven-databases-in-seven-weeks http://pragprog.com/book/rwdata/seven-databases-in-seven-wee... and grep for the phrase "polyglot persistence" you'll find some great information to get you started. a video promo for the book: https://www.youtube.com/watch?v=bSAc56YCOaE https://www.youtube.com/watch?v=bSAc56YCOaE
  - twic 13y ago
    Can you talk about what sort of queries you do in Neo4j? Graphs are perhaps the only area where NoSQL may beat relational databases for expressiveness, rather than alleged scalability or ease of use. I've read the documentation for Neo4j and Cypher, but never used them in anger, so i'd be really interested to know more about how they are actually used. In particular, i would love to be able to make a comparison between a graph database and a relational database queried using recursive common table expressions, in the context of a real-world problem.
    - h1karu 13y ago
      The approach I like is to use a key/value document store as the system of record (SOR). If any data conflicts arise between Redis, the document store, and the graph database, then the SOR wins. A good SOR should contain all of the data necessary to rebuild any other data source in its domain. Graph databases can scale write throughput fine when those writes are being fed to it from a single source so it's best to have a service who's sole purpose is to keep the graph database in sync with the SOR. The graph should only store the data you actually need in order to get the queries that you want. The specific queries in my domain are exactly the kind of thing I'm not supposed to talk about, but I will say you can do things like: Give me 50 users who who live near me (lat/long bounding box), who I'm not already following, who I have not already sent a message to, who have at least two friends who also live near me, and who have matching tags, order by number of followers. there are a bunch of great examples in this awesome book: http://www.amazon.com/Graph-Databases-Ian-Robinson/dp/1449356265 http://www.amazon.com/Graph-Databases-Ian-Robinson/dp/144935...