Apache Samza uses RocksDB as the storage engine for local storage. This allows for stateful stream processing and here's a very good overview.
My use case:
I have multiple streams of events that I wish to process taken from a system such as Apache Kafka.
These events create state - the state I wish to track is based on previous messages received.
I wish to generate new stream events based on the calculated state.
The input stream events are highly connected and a graph such as OrientDB / Neo4J is the ideal medium for querying the data to create the new stream events.
My question:
Is it possible to use a non-KV store as the local storage for Samza? Has anyone ever done this with OrientDB / Neo4J and is anyone aware of an example?
I've been evaluating Samza and I'm by no means an expert, but I'd recommend you to read the official documentation, and even read through the source code—other than the fact that it's in Scala, it's remarkably approachable.
In this particular case, toward the bottom of the documentation's page on State Management you have this:
Other storage engines
Samza’s fault-tolerance mechanism (sending a local store’s writes to a replicated changelog) is completely decoupled from the storage engine’s data structures and query APIs. While a key-value storage engine is good for general-purpose processing, you can easily add your own storage engines for other types of queries by implementing the StorageEngine interface. Samza’s model is especially amenable to embedded storage engines, which run as a library in the same process as the stream task.
Some ideas for other storage engines that could be useful: a persistent heap (for running top-N queries), approximate algorithms such as bloom filters and hyperloglog, or full-text indexes such as Lucene. (Patches accepted!)
I actually read through the code for the default StorageEngine implementation about two weeks ago to gain a better sense of how it works. I definitely don't know enough to say much intelligently about it, but I can point you at it:
https://github.com/apache/samza/tree/master/samza-kv-rocksdb/src/main/scala/org/apache/samza/storage/kv
https://github.com/apache/samza/tree/master/samza-kv/src/main/scala/org/apache/samza/storage/kv
The major implementation concerns seem to be:
Logging all changes to a topic so that the store's state can be restored if a task fails.
Restoring the store's state in a performant manner
Batching writes and caching frequent reads in order to save on trips to the raw store.
Reporting metrics about the use of the store.
Do the input stream events define one global graph, or multiple graphs for each matching Kafka/Samza partition? That is important as Samza state is local not global.
If it's one global graph, you can update/query a separate graph system from the Samza task process method. Titan on Cassandra would one such graph system.
If it's multiple separate graphs, you can use the current RocksDB KV store to mimic graph database operations. Titan on Cassandra does just that - uses Cassandra KV store to store and query the graph. Graphs are stored either via matrix (set [i,j] to 1 if connected) or edge list. For each node, use it as the key and store its set of neighbors as the value.
Related
I was going through the article, https://learn.microsoft.com/en-us/azure/architecture/patterns/cqrs which says, "If separate read and write databases are used, they must be kept in sync". One obvious benefit I can understand from having separate read replicas is that they can be scaled horizontally. However, I have some doubts:
It says, "Updating the database and publishing the event must occur in a single transaction". My understanding is that there is no guarantee that the updated data will be available immediately on the read-only nodes because it depends on when the event will be consumed by the read-only nodes. Did I get it correctly?
Data must be first written to read-only nodes before it can be read i.e. write operations are also performed on the read-only nodes. Why are they called read-only nodes? Is it because the write operations are performed on these nodes not directly by the data producer application; but rather by some serverless function (e.g. AWS Lambda or Azure Function) that picks up the event from the topic (e.g. Kafka topic) to which the write-only node has sent the event?
Is the data sharded across the read-only nodes or does every read-only node have the complete set of data?
All of these have "it depends"-like answers...
Yes, usually, although some implementations might choose to (try to) update read models transactionally with the update. With multiple nodes you're quickly forced to learn the CAP theorem, though, and so in many CQRS contexts, eventual consistency is just accepted as a feature, as the gains from tolerating it usually significantly outweigh the losses.
I suspect the bit you quoted anyway refers to transactionally updating the write store with publishing the event. Even this can be difficult to achieve, and is one of the problems event sourcing seeks to solve.
Yes. It's trivially obvious - in this context - that data must be written before it can be read, but your apps as consumers of the data see them as read-only.
Both are valid outcomes. Usually this part is less an application concern and is more delegated to the capabilities of your chosen read-model infrastructure (Mongo, Cosmos, Dynamo, etc).
If I have a use case where I need to join two streams or aggregate some kind of metrics from a single stream, and I use keyed streams to partition the events, how does Flink handle the operations for hot partitions where the data might not fit into memory and needs to be split across partitions?
Flink doesn't do anything automatic regarding hot partitions.
If you have a consistently hot partition, you can manually split it and pre-aggregate the splits.
If your concern is about avoiding out-of-memory errors due to unexpected load spikes for one partition, you can use a state backend that spills to disk.
If you want more dynamic data routing / partitioning, look at the Stateful Functions API or the Dynamic Data Routing section of this blog post.
If you want auto-scaling, see Autoscaling Apache Flink with Ververica Platform Autopilot.
I'm researching graph databases for a work project. Since our data is highly connected, it appears that a graph database would be a good option for us.
One of the first graph DB options I've run into is neo4j, and for the most part, I like it. However, I have one question about neo4j to which I cannot find the answer: Can I get neo4j to store the entire graph in-memory? If so, how does one configure this?
The application I'm designing needs to be lightning-fast. I can't afford to wait for the db to go to disk to retrieve the data I'm searching for. I need the entire DB to be held in-memory to reduce the query time.
Is there a way to hold the entire neo4j DB in-memory?
Thanks!
Further to Bruno Peres' answer, if you want to run a regular server instance, Neo4j will load the entire graph into memory when resources are sufficient. This does indeed improve performance.
The Manual has a chapter on configuring memory.
The page cache portion holds graph data and indexes - this is configured via the dbms.memory.pagecache.size property in neo4j.conf. If it is large enough, the whole graph will be stored in memory.
The heap space portion is for query execution, state management, etc. This is set via the dbms.memory.heap.initial_size and
dbms.memory.heap.max_size properties. Generally these two properties should be set to the same value, so that the whole heap is allocated on startup.
If the sole purpose of the server is to run Neo4j, you can allocate most of the memory to the heap and page cache, leaving enough left over for operating system tasks.
Holding Very Large Graphs In Memory
At Graph Connect in San Francisco, 2016, Neo4j's CTO, Jim Webber, in his typical entertaining fashion, gave details on servers that have a very large amount of high performance memory - capable of holding an entire large graph in memory. He seemed suitably impressed by them. I forget the name of the machines, but if you're interested, the video archive should have details.
Neo4j isn't designed to hold the entire graph in main memory. This leaves you with a couple of options. You can either play around with the config parameters (as Jasper Blues already explained in more details) OR you can configure Neo4j to use RAMDisk.
The first option probably won't give you the best performance as only the cache is held in memory.
The challenge with the second approach is that everything is in-memory which means that the system isn't durable and the writes are inefficient.
You can take a look at Memgraph (DISCLAIMER: I'm the co-founder and CTO). Memgraph is a high-performance, in-memory transactional graph database and it's openCypher and Bolt compatible. The data is first stored in main memory before being written to disk. In other words, you can choose to make a tradeoff between write speed and safety.
I've been reading Neo4j's Operational Manual on Cache Sharding, and posts all over the web, however I can hardly find any detailed example on how to configure HAProxy for cache sharding(yes the one on Operation Manual is rather brief) on a real-world graph, which may contain multiple node labels.
Has anyone ever done this before? Would be lovely if you could share your experience.
Moreover, I'm a bit confused on the mechanism of the way to shard the graph using HAProxy. How do sub-graphs get cached on certain slaves, merely by providing rules in HAProxy? It surprised me to learn that cache sharding isn't handled by Neo4j.
The goal is to send queries hitting the same region of your graph always to the same instance. This of course means that the request data indicates the region. What to use as "region indicator" is heavily depending on the structure and shape of your graph.
In a lot of cases of customer facing applications people successfully used the current user id and set it as additional http header which is then evaluated by haproxy.
I have streaming data coming into my consumer app that I ultimately want to show up in Hive/Impala. One way would be to use Hive based APIs to insert the updates in batches to the Hive Table.
The alternate approach is to write the data directly into HDFS as a avro/parquet file and let hive detect the new data and suck it in.
I tried both approaches in my dev environment and the 'only' drawback I noticed was high latency writing to hive and/or failure conditions I need to account for in my code.
Is there an architectural design pattern/best practices to follow?