We are implementing few business flows in financial area. The requirement (unfortunately, not very specific) from the regulatory is to have a data lineage for auditing purpose.
The flow contains 2 parts: synchronous and asynchronous. The syncronous part is a payment attempt containing bunch of info about point of sale, the customer and the goods. The asynchronous part is a batch process that feeds the credit assessment data model with a newly-calculated portion of variables on an hourly basis. The variables might include some aggregations like balances and links to historical transactions.
For calculating the asynchronous part we ingest the data from multiple relational DBs and store them in HDFS in a raw format (rows from tables in csv format).
When storing the data on HDFS is done a job based on Spring XD that calculates some aggregations and produces the data for the synchronous part is triggered.
We have relational data, raw data on HDFS and MapReduce jobs relying on POJOs that describe the relevant semantics and the transformations implemented in SpringXD.
So, the question is how to handle the auditing in the scenario described above?
We need at any point in time to be able to explain why a specific decision was made and also be able to explain how each and every variable used in the policy (synchronous or near-real-time flow) was calculated.
I looked at existing Hadoop stack and it looks like currently no tool could provide with a good enterprise-ready auditing capabilities.
My thinking is to start with custome implementation that includes>
A business glossary with all the business terms
Operational and technical metadata - logging transformation execution for each entry into a separate store.
log changes to a business logic (use data from version control where the business rules and transformations are kept).
Any advice or sharing your experience would be greatly appreciated!
Currently Cloudera sets the industry standard for Data Lineage/Data Governance in the big data space.
Glossary, metadata and historically run (versions of) queries can all be facilitated.
I do realize some of this may not have been in place when you asked the question, but it certainly is now.
Disclaimer: I am an employee of Cloudera
Related
Background information
We sell an API to users, that analyzes and presents corporate financial-portfolio data derived from public records.
We have an "analytical data warehouse" that contains all the raw data used to calculate the financial portfolios. This data warehouse is fed by an ETL pipeline, and so isn't "owned" by our API server per se. (E.g. the API server only has read-only permissions to the analytical data warehouse; the schema migrations for the data in the data warehouse live alongside the ETL pipeline rather than alongside the API server; etc.)
We also have a small document store (actually a Redis instance with persistence configured) that is owned by the API layer. The API layer runs various jobs to write into this store, and then queries data back as needed. You can think of this store as a shared persistent cache of various bits of the API layer's in-memory state. The API layer stores things like API-key blacklists in here.
Problem statement
All our input data is denominated in USD, and our calculations occur in USD. However, we give our customers the query-time option to convert the response just-in-time to another currency. We do this by having the API layer run a background job to scrape exchange-rate data, and then cache it in the document store. Individual API-layer nodes then do (in-memory-cached-with-TTL) fetches from this exchange-rates key in the store, whenever a query result needs to be translated into a specific currency.
At first, we thought that this unit conversion wasn't really "about" our data, just about the API's UX, and so we thought this was entirely an API-layer concern, where it made sense to store the exchange-rates data into our document store.
(Also, we noticed that, by not pre-converting our DB results into a specific currency on the DB side, the calculated results of a query for a particular portfolio became more cache-friendly; the way we're doing things, we can cache and reuse the portfolio query results between queries, even if the queries want the results in different currencies.)
But recently we've been expanding into also allowing partner clients to also execute complex data-science/Business Intelligence queries directly against our analytical data warehouse. And it turns out that they will also, often, need to do final exchange-rate conversions in their BI queries as well—despite there being no API layer involved here.
It seems like, to serve the needs of BI querying, the exchange-rate data "should" actually live in the analytical data warehouse alongside the financial data; and the ETL pipeline "should" be responsible for doing the API scraping required to fetch and feed in the exchange-rate data.
But this feels wrong: the exchange-rate data has a different lifecycle and integrity constraints than our financial data. The exchange rates are dirty and ephemeral point-in-time samples attained by scraping, whereas the financial data is a reliable historical event stream. The exchange rates get constantly updated/overwritten, while the financial data is append-only. Etc.
What is the best practice for serving the needs of analytical queries that need to access backend "application state" for "query result presentation" needs like this? Or am I wrong in thinking of this exchange-rate data as "application state" in the first place?
What I find interesting about your scenario is about when the exchange rate data is applicable.
In the case of the API, it's all about the realtime value in the other currency and it makes sense to have the most recent value in your API app scope (Redis).
However, I assume your analytical data warehouse has tables with purchases that were made at a certain time. In those cases, the current exchange rate is not really relevant to the value of the transaction.
This might mean that you want to store the exchange rate history in your warehouse or expand the "purchases" table to store the values in all the currencies at that moment.
In an application we have to send sensory data stream from multiple clients to a central server over internet. One obvious solution is to use MOMs (Message Oriented Middlewares) such as Kafka, but I recently learned that we can do this with data base synchronization tools such as oracle Materialized View.
The later approach works in some application (sending data from a central server to multiple clients, inverse directin of our application), but what is the pros and cons of it in our application? Which one is better for sending sensory data stream from multiple (~100) clients to server in terms of speed, security, etc.?
Thanks.
P.S.
For more detail consider an application in which many (about 100) clients have to send streaming data (1MB data per minute) to a central server over internet. The data are needed in server for the sake of online monitoring, analysis and some computation such as machine learning and data mining tasks.
My question is about the difference between db-to-db connection and streaming solutions such as kafka for trasfering data from clients to server.
Prologue
I'm going to try and break your question down into in order to get a clearer understanding of your current requirements and then build it back up again. This has taken a long time to write so I'd really appreciate it if you do two things off the back of it:
Be sceptical - there's absolutely no substitute for testing things yourself. The internet is very useful as a guide but there's no guarantee that the help you receive (if this answer is even helpful!) is the best thing for your specific situation. It's impossible to completely describe your current situation in the space allotted and so any answer is, of necessity, going to be lacking somewhere.
Look again at how you explained yourself - this is a valid question that's been partially stopped by a lack of clarity in your description of the system and what you're trying to achieve. Getting someone unfamiliar with your system to look over your question before posting a complex question may help.
Problem definition
sensory data stream from multiple clients to a central server
You're sending data from multiple locations to a single persistence store
online monitoring
You're going to be triggering further actions based off the raw data and potentially some aggregated data
analysis and some computation such as machine learning and data mining tasks
You're going to be performing some aggregations on the clients' data, i.e. you require aggregations of all of the clients' data to be persisted (however temporarily) somewhere
Further assumptions
Because you're talking about materialized views we can assume that all the clients persist data in a database, probably Oracle.
The data coming in from your clients is about the same topic.
You've got ~100 clients, at that amount we can assume that:
the number of clients might change
you want to be able to add clients without increasing the number of methods of accessing data
You don't work for one of Google, Amazon, Facebook, Quantcast, Apple etc.
Architecture diagram
Here, I'm not making any comment on how it's actually going to work - it's the start of a discussion based on my lack of knowledge of your systems. The "raw data persistence" can be files, Kafka, a database etc. This is description of the components that are going to be required and a rough guess as to how they will have to connect.
Applying assumed architecture to materialized views
Materialized views are a persisted query. Therefore you have two choices:
Create a query that unions all 100 clients data together. If you add or remove a client you must change the query. If a network issue occurs at any one of your clients then everything fails
Write and maintain 100 materialized views. The Oracle database at your central location has 100 incoming connections.
As you can probably guess from the tradeoffs you'll have to make I do not like materialized views as the sole solution. We should be trying to reduce the amount of repeated code and single points of failure.
You can still use materialized views though. If we take our diagram and remove all the duplicated arrows in your central location it implies two things.
There is a single service that accepts incoming data
There is a single service that puts all the incoming data into a single place
You could then use a single materialized view for your aggregation layer (if your raw data persistence isn't in Oracle you'll first have to put the data into Oracle).
Consequences of changes
Now we've decided that you have a single data pipeline your decisions actually become harder. We've decoupled your clients from the central location and the aggregation layer from our raw data persistence. This means that the choices are now yours but they're also considerably easier to change.
Reimagining architecture
Here we need to work out what technologies aren't going to change.
Oracle databases are expensive and you're pushing 140GB/day into yours (that's 50TB/year by the way, quite a bit). I don't know if you're actually storing all the raw data but at those volumes it's less likely that you are - you're only storing the aggregations
I'm assuming you've got some preferred technologies where your machine learning and data mining happen. If you don't then consider getting some to prevent madness supporting everything
Putting all of this together we end up with the following. There's actually only one question that matters:
How many times do you want to read your raw data off your database.
If the answer to that is once then we've just described middleware of some description. If the answer is more than once then I would reconsider unless you've got some very good disks. Whether you use Kafka for this middle layer is completely up to you. Use whatever you're most familiar with and whatever you're most willing to invest the time into learning and supporting. The amount of data you're dealing with is non-trivial and there's going to be some trial and error getting this right.
One final point about this; we've defined a data pipeline. A single method of data flowing through your system. In doing so, we've increased the flexibility of the system. Want to add more clients, no need to do anything. Want to change the technology behind part of the system, as long as the interface remains the same there's no issue. Want to send data elsewhere, no problem, it's all in the raw data persistence layer.
I would like to store a large amount of timeseries from devices. Also these timeseries have to be validated, can be modified by an operator and have to be exported to other systems. Holes in the timeseries must be found. Timeseries must be shown in the UI filtered by serialnumber and date range.
We have thought about using hadoop, hbase, opentsdb and spark for this scenario.
What do you think about it? Can Spark connect to opentsdb easily?
Thanks
OpenTSDB is really great for storing large amount of time series data. Internally, it is underpinned by HBase - which means that it had to find a way around HBase's limitations in order to perform well. As a result, the representation of time series is highly optimized and not easy to decode. AFAIK, there is no out-of-the-box connector that would allow to fetch data from OpenTSDB into Spark.
The following GitHub project might provide you with some guidance:
Achak1987's connector
If you are looking for libs that would help you with time series, have a look at spark-ts - it contains useful functions for missing data imputation as well.
Warp 10 offers the WarpScript language which can be used from Spark/Pig/Flink to manipulate time series and access data stored in Warp 10 via a Warp10InputFormat.
Warp 10 is Open Source and available at www.warp10.io
Disclaimer: I'm CTO of Cityzen Data, maker of Warp 10.
Take a look at Axibase Time Series Database which has a rather unique versioning feature to maintain a history of value changes for the same timestamp. Once enabled with per-metric granularity, the database keeps track of source, status and times of value modifications for audit trail or data reconciliation.
We have customers streaming data from Spark apps using Network API, typically once data is enriched with additional metadata (aks series tags) for downstream reporting.
You can query data from ATSD with REST API or SQL.
Disclaimer: I work for Axibase.
I am going through the lambda architecture and understanding how it can be used to build fault tolerant big data systems.
I am wondering how batch layer is useful when everything can be stored in realtime view and generate the results out of it? is it because realtime storage cant be used to store all of the data, then it wont be realtime as the time taken to retrieve the data is dependent on the the space it took for the data to store.
Why batch layer
To save Time and Money!
It basically has two functionalities,
To manage the master dataset (assumed to be immutable)
To pre-compute the batch views for ad-hoc querying
Everything can be stored in realtime view and generate the results out of it - NOT TRUE
The above is certainly possible, but not feasible as data could be 100's..1000's of petabytes and generating results could take time.. a lot of time!
Key here, is to attain low-latency queries over large dataset. Batch layer is used for creating batch views (queries served with low-latency) and realtime layer is used for recent/updated data which is usually small. Now, any ad-hoc query can be answered by merging results from batch views and real-time views instead of computing over all the master dataset.
Also, think of a query (same query?) running again and again over huge dataset.. loss of time and money!
Further to the answer provided by #karthik manchala, data Processing can be handled in three ways - Batch, Interactive and Real-time / Streaming.
I believe, your reference to real-time is more with interactive response than to streaming as not all use cases are streaming related.
Interactive responses are where the response can be expected anywhere from sub-second to few seconds to minutes, depending on the use case. Key here is to understand that processing is done on data at rest i.e. already stored on a storage medium. User interacts with the system while processing and hence waits for the response. All the efforts of Hive on Tez, Impala, Spark core etc are to address this issue and make the responses as fast as possible.
Streaming on the other side is where data streams into the system in real-time - for example twitter feeds, click streams etc and processing need to be done as soon as the data is generated. Frameworks like Storm, Spark Streaming address this space.
The case for batch processing is to address scenarios where some heavy-lifting need to be done on a huge dataset before hand such that user would be made believe that the responses he sees are real-time. For example, indexing a huge collection of documents into Apache Solr is a batch job, where indexing would run for minutes or possibly hours depending on the dataset. However, user who queries the Solr index would get the response in sub-second latency. As you can see, indexing cannot be achieved in real-time as there may be hue amounts of data. Same is the case with Google search, where indexing would be done in a batch mode and the results are presented in interactive mode.
All the three modes of data processing are likely involved in any organisation grappling with data challenges. Lambda Architecture addresses this challenge effectively to use the same data sources for multiple data processing requirements
You can check out the Kappa-Architecture where there is no seperate Batch-Layer.
Everything is analyzed in the Stream-Layer. You can use Kafka in the right configuration as as master-datasetstorage and save computed data in a database as your view.
If you want to recompute, you can start a new Stream-Processing job and recompute your view from Kafka into your database and replace your old view.
It is possible to use only the Realtime view as the main storage for adhoc query but as it is already mentioned in other answers, it is faster if you have much data to do batch-processing and stream-processing seperate instead of doing batch-jobs as a stream-job. It depends on the size of your data.
Also it is cheaper to have a storage like hdfs instead of a database for batch-computing.
And the last point in many cases you have different algorithms for batch and stream processing, so you need to do it seperate. But basically it is possible to only use the "realtime view" as your batch-and stream-layer also without using Kafka as masterset. It depends on your usecase.
I am writing an ETL (in python with a mongodb backend) and was wondering : what kind of standard functions and tools an ETL should have to be called an ETL ?
This ETL will be as general purpose as possible, with a scriptable and modular approach. Mostly it will be used to keep different databases in sync, and to import/export datasets in different formats (xml and csv) I don't need any multidimensional tools, but it is a possibility that it'll needed later.
Let's think of the ETL use cases for a moment.
Extract.
Read databases through a generic DB-API adapter.
Read flat files through a similar adapter.
Read spreadsheets through a similar adapter.
Cleanse.
Arbitrary rules
Filter and reject
Replace
Add columns of data
Profile Data.
Statistical frequency tables.
Transform (see cleanse, they're two use cases with the same implementation)
Do dimensional conformance lookups.
Replace values, or add values.
Aggregate.
At any point in the pipeline
Load.
Or prepare a flat-file and run the DB product's loader.
Further, there are some additional requirements that aren't single use cases.
Each individual operation has to be a separate process that can be connected in a Unix pipeline, with individual records flowing from process to process. This uses all the CPU resources.
You need some kind of time-based scheduler for places that have trouble reasoning out their ETL preconditions.
You need an event-based schedule for places that can figure out the preconditions for ETL processing steps.
Note. Since ETL is I/O bound, multiple threads does you little good. Since each process runs for a long time -- especially if you have thousands of rows of data to process -- the overhead of "heavyweight" processes doesn't hurt.
Here's a random list, in no particular order:
Connect to a wide range of sources, including all the major relational databases.
Handle non-relational data sources like text files, Excel, XML, etc.
Allow multiple sources to be mapped into a single target.
Provide a tool to help map from source to target fields.
Offer a framework for injecting transformations at will.
Programmable API for writing complex transformations.
Optimize load process for speed.
Automatic / heuristic mapping of column names. E.g simple string mappings:
DB1: customerId
DB2: customer_id
I find a lot of the work I (have) done in DTS / SSIS could've been automatically generated.
not necessarily "required functionality", but would keep a lot of your users very happy indeed.