I was wondering why Dataflow does not support 'SortByKey' like Apache Spark.
I have a huge table in BigQuery that I cannot sort it because "Order By" is not scalable. So, I was thinking to move the output of BigQuery to Dataflow and sort it there. But, there is no SortByKey and it seems I have to write a combiner.
Any suggestions will be appreciated.
Sorting (especially by key) requires globally serial processing, which is not a scalable operation. Apache Beam / Dataflow does not provide such support, as it is frequently unnecessary.
There are a variety of alternatives that generally address the need more scalably. For instance, you can sort the values within each key, which allows each key to be processed in parallel. Another common use case is TopN either globally or per-key. Again, this can be supported much more efficiently than actually sorting.
Could you elaborate on what you need to sort by and why? It would make it possible to identify options for implementing this within the Beam and Dataflow SDKs.
Coming from an SQL guy, I'm looking to use NoSQL in production. One thing I noticed is that transferring from one provider to another is not going to be as easy as importing/exporting in "normal" SQL. Correct me if I'm wrong but it appears each flavor of NoSQL has its own "scheme".
My question is, say I chose Google Datastore today, then some time in the future I decide to move my data to Amazon DynamoDB or some hosted MongoDB service, for whatever reason it may be (price/performance/etc).. Do I need to code my own transition script or is there a standard way/tool to move across different NoSQL solutions (like simple import/export in traditional SQL DBs)?
Yes, you guessed right. The databases you listed are all wildly different. Consequently, there is no automated way of moving data between them (that I heard of). Nor is there any sense.
MongoDB, for example, supports quite rich set of operations, compared to which the key-value API of DynamoDB looks primitive. So unless your app only uses "get_item/put_item" operations, you can't really switch between different NoSQL databases.
Apologies if the title made no absolute sense. But, on the other hand, I would like to know if there is any programming model which would let us use Infinispan cache as a real datastore and not just a grid on top of an underlying rdbms.
I know Key-Value stores have real limitations but I couldn't stop thinking about the possibilities of an in-memory solution with all or a subset of RDBMS functionalities. For example: If I want to retrieve a particular set of Keys based on value>34.56%, just like how we would use a where clause in an sql stmt.
My doubt is not specific to infinispan but any IMKVS for that purpose. I know it's a shot in the dark considering the data structures and algorithms behind IMKVS specifications.
Any help or direction to resources which talk about these lines would be of great help.
I suggest you write down all the queries that you execute against SQL DB and check if these could be translated into KVS language.
In Infinispan you can index the values and execute queries for such filtering, but you can't do any table joins.
If you are in need for more powerful API, specifically using JPA, take a look at Hibernate OGM.
And while KVSs offer some level of reliability, in practice I wouldn't trust the documentation too much. You need to perform extensive testing of your system and check that you can retrieve the data even after various types of crashes and network failures (or that you can live with throwing the data away).
I have few Gigabytes text file in format:
{"user_ip":"x.x.x.x", "action_type":"xxx", "action_data":{"some_key":"some_value"...},...}
each entry is one line.
First I would like to easily find entries for given ip. This part is easy because I can use grep for example. However even for this I would like to find better solution because I would like to get response as fast as possible.
Next part is more complicated because I would like to find entries from selected ip and of selected type and with particular value of some_key in action_data.
Probably I would have to convert this file to SQL db (probably SQLite, because it will be desktop APP), but I would ask if there are exists better solutions?
You could take a look at MongoDB, a document based database. With it you essentially store JSON objects that you can then index and easily query in an efficient way. You can find about how to query in the docs: Querying.
Yes, put it into a database, any database. Then querying it will be straightforward.
Just wanted to mention that Oracle Berkeley DB 11gR2 (released on April 1st, 2010) introduces support for a SQL API. In fact, the SQL API is the sqlite3() API. So, as Jason mentioned, if you'd like the ease-of-use of SQLite, combined with the scalability and concurrency of Berkeley DB, you can now get both things in a single library.
Regards,
Dave
If you need the relational guarantees of an SQL-based DB, definitely go ahead with SQLite. It will allow for fast queries, joins, aggregations, sorts, and overall any sort of search you could possibly dream up. It sounds like this is just a big list of Actions performed by users at some IP, so you'll probably want to use some sort of sequence as your primary key since none of the other attributes look like good candidates.
On the other hand, if you just need to do very simple queries, e.g. look up entries by IP, look up entries by action type, etc., you might want to look into Oracle Berkeley DB. As long as you don't need any searches that are too fancy, Berkeley DB will let you store Terabytes of data and access them at record speed.
So look over both and see what's best for your use case. They're good for different things, which might be why both are available as storage systems on Android, for instance. I think SQLite will probably win out, but when thinking about embedded local DB systems you should always at least consider both of these technologies.
Can you please point to alternative data storage tools and give good reasons to use them instead of good-old relational databases? In my opinion, most applications rarely use the full power of SQL--it would be interesting to see how to build an SQL-free application.
Plain text files in a filesystem
Very simple to create and edit
Easy for users to manipulate with simple tools (i.e. text editors, grep etc)
Efficient storage of binary documents
XML or JSON files on disk
As above, but with a bit more ability to validate the structure.
Spreadsheet / CSV file
Very easy model for business users to understand
Subversion (or similar disk based version control system)
Very good support for versioning of data
Berkeley DB (Basically, a disk based hashtable)
Very simple conceptually (just un-typed key/value)
Quite fast
No administration overhead
Supports transactions I believe
Amazon's Simple DB
Much like Berkeley DB I believe, but hosted
Google's App Engine Datastore
Hosted and highly scalable
Per document key-value storage (i.e. flexible data model)
CouchDB
Document focus
Simple storage of semi-structured / document based data
Native language collections (stored in memory or serialised on disk)
Very tight language integration
Custom (hand-written) storage engine
Potentially very high performance in required uses cases
I can't claim to know anything much about them, but you might also like to look into object database systems.
Matt Sheppard's answer is great (mod up), but I would take account these factors when thinking about a spindle:
Structure : does it obviously break into pieces, or are you making tradeoffs?
Usage : how will the data be analyzed/retrieved/grokked?
Lifetime : how long is the data useful?
Size : how much data is there?
One particular advantage of CSV files over RDBMSes is that they can be easy to condense and move around to practically any other machine. We do large data transfers, and everything's simple enough we just use one big CSV file, and easy to script using tools like rsync. To reduce repetition on big CSV files, you could use something like YAML. I'm not sure I'd store anything like JSON or XML, unless you had significant relationship requirements.
As far as not-mentioned alternatives, don't discount Hadoop, which is an open source implementation of MapReduce. This should work well if you have a TON of loosely structured data that needs to be analyzed, and you want to be in a scenario where you can just add 10 more machines to handle data processing.
For example, I started trying to analyze performance that was essentially all timing numbers of different functions logged across around 20 machines. After trying to stick everything in a RDBMS, I realized that I really don't need to query the data again once I've aggregated it. And, it's only useful in it's aggregated format to me. So, I keep the log files around, compressed, and then leave the aggregated data in a DB.
Note I'm more used to thinking with "big" sizes.
The filesystem's prety handy for storing binary data, which never works amazingly well in relational databases.
Try Prevayler:
http://www.prevayler.org/wiki/
Prevayler is alternative to RDBMS. In the site have more info.
If you don't need ACID, you probably don't need the overhead of an RDBMS. So, determine whether you need that first. Most of the non-RDBMS answers provided here do not provide ACID.
Custom (hand-written) storage engine / Potentially very high performance in required uses cases
http://www.hdfgroup.org/
If you have enormous data sets, instead of rolling your own, you might use HDF, the Hierarchical Data Format.
http://en.wikipedia.org/wiki/Hierarchical_Data_Format:
HDF supports several different data models, including multidimensional arrays, raster images, and tables.
It's also hierarchical like a file system, but the data is stored in one magic binary file.
HDF5 is a suite that makes possible the management of extremely large and complex data collections.
Think petabytes of NASA/JPL remote sensing data.
G'day,
One case that I can think of is when the data you are modelling cannot be easily represented in a relational database.
Once such example is the database used by mobile phone operators to monitor and control base stations for mobile telephone networks.
I almost all of these cases, an OO DB is used, either a commercial product or a self-rolled system that allows heirarchies of objects.
I've worked on a 3G monitoring application for a large company who will remain nameless, but whose logo is a red wine stain (-: , and they used such an OO DB to keep track of all the various attributes for individual cells within the network.
Interrogation of such DBs is done using proprietary techniques that are, usually, completely free from SQL.
HTH.
cheers,
Rob
Object databases are not relational databases. They can be really handy if you just want to stuff some objects in a database. They also support versioning and modify classes for objects that already exist in the database. db4o is the first one that comes to mind.
In some cases (financial market data and process control for example) you might need to use a real-time database rather than a RDBMS. See wiki link
There was a RAD tool called JADE written a few years ago that has a built-in OODBMS. Earlier incarnations of the DB engine also supported Digitalk Smalltalk. If you want to sample application building using a non-RDBMS paradigm this might be a start.
Other OODBMS products include Objectivity, GemStone (You will need to get VisualWorks Smalltalk to run the Smalltalk version but there is also a java version). There were also some open-source research projects in this space - EXODUS and its descendent SHORE come to mind.
Sadly, the concept seemed to die a death, probably due to the lack of a clearly visible standard and relatively poor ad-hoc query capability relative to SQL-based RDMBS systems.
An OODBMS is most suitable for applications with core data structures that are best represented as a graph of interconnected nodes. I used to say that the quintessential OODBMS application was a Multi-User Dungeon (MUD) where rooms would contain players' avatars and other objects.
You can go a long way just using files stored in the file system. RDBMSs are getting better at handling blobs, but this can be a natural way to handle image data and the like, particularly if the queries are simple (enumerating and selecting individual items.)
Other things that don't fit very well in a RDBMS are hierarchical data structures and I'm guessing geospatial data and 3D models aren't that easy to work with either.
Services like Amazon S3 provide simpler storage models (key->value) that don't support SQL. Scalability is the key there.
Excel files can be useful too, particularly if users need to be able to manipulate the data in a familiar environment and building a full application to do that isn't feasible.
There are a large number of ways to store data - even "relational databse" covers a range of alternatives from a simple library of code that manipulates a local file (or files) as if it were a relational database on a single user basis, through file based systems than can handle multiple-users to a generous selection of serious "server" based systems.
We use XML files a lot - you get well structured data, nice tools for querying same the ability to do edits if appropriate, something that's human readable and you don't then have to worry about the db engine working (or the workings of the db engine). This works well for stuff that's essentially read only (in our case more often than not generated from a db elsewhere) and also for single user systems where you can just load the data in and save it out as required - but you're creating opportunities for problems if you want multi-user editing - at least of a single file.
For us that's about it - we're either going to use something that will do SQL (MS offer a set of tools that run from a .DLL to do single user stuff all the way through to enterprise server and they all speak the same SQL (with limitations at the lower end)) or we're going to use XML as a format because (for us) the verbosity is seldom an issue.
We don't currently have to manipulate binary data in our apps so that question doesn't arise.
Murph
One might want to consider the use of an LDAP server in the place of a traditional SQL database if the application data is heavily key/value oriented and hierarchical in nature.
BTree files are often much faster than relational databases. SQLite contains within it a BTree library which is in the public domain (as in genuinely 'public domain', not using the term loosely).
Frankly though, if I wanted a multi-user system I would need a lot of persuading not to use a decent server relational database.
Full-text databases, which can be queried with proximity operators such as "within 10 words of," etc.
Relational databases are an ideal business tool for many purposes - easy enough to understand and design, fast enough, adequate even when they aren't designed and optimized by a genius who could "use the full power," etc.
But some business purposes require full-text indexing, which relational engines either don't provide or tack on as an afterthought. In particular, the legal and medical fields have large swaths of unstructured text to store and wade through.
Also:
* Embedded scenarios - Where usually it is required to use something smaller then a full fledged RDBMS. Db4o is an ODB that can be easily used in such case.
* Rapid or proof-of-concept development - where you wish to focus on the business and not worry about persistence layer
CAP theorem explains it succinctly. SQL mainly provides "Strong Consistency: all clients see the same view, even in presence of updates".
K.I.S.S: Keep It Small and Simple
I would offer RDBMS :)
If you do not wont to have troubles with set up/administration go for SQLite.
Built in RDBMS with full SQL support. It even allows you to store any type of data in any column.
Main advantage against for example log file: If you have huge one, how are you going to search in it? With SQL engine you just create index and speed up operation dramatically.
About full text search: SQLite has modules for full text search too..
Just enjoy nice standard interface to your data :)
One good reason not to use a relational database would be when you have a massive data set and want to do massively parallel and distributed processing on the data. The Google web index would be a perfect example of such a case.
Hadoop also has an implementation of the Google File System called the Hadoop Distributed File System.
I would strongly recommend Lua as an alternative to SQLite-kind of data storage.
Because:
The language was designed as a data description language to begin with
The syntax is human readable (XML is not)
One can compile Lua chunks to binary, for added performance
This is the "native language collection" option of the accepted answer. If you're using C/C++ as the application level, it is perfectly reasonable to throw in the Lua engine (100kB of binary) just for the sake of reading configs/data or writing them out.