Hadoop “fails us”
datanami.com
datanami.com
You are making a very common mistake of coupling the file system and the map reduce implementation and the scheduler . You are right about this post and kafka though. Let me expand on this point a bit more.
Hadoop isn't what it was in 2004. It's now a complex beast with several decoupled components which now make it very hard to identify what "hadoop" is for people outside the space.
The hadoop ecosystem is actually very healthy if you actually look at all the streaming platforms built on top of it (kafka, apex, spark, flink,tez,..)
There are also databases such as hbase,cassandra and more recently kudu, specialized for different workloads. Don't even get me started on all the sql implementations (again with their own trade offs) such as impala and hive.
If we step back for a second here and focus on just the compute part: Yes map reduce is for the most part dead. This is supplanted by the streaming and batch platforms such as flink and spark.
The scheduler part (YARN) is competing with mesos largely in part thanks to spark and flink being able to leverage both with mesos being way more flexible. (Most hadoop distros only use YARN though).
Then we also have the distributed consensus part in zookeeper. Etcd is up and coming in this piece but your hadoop cluster uses zookeeper (both mesos and kafka rely on it for example)
For what it's worth, all three of them seemed pretty down on Hadoop.
A quick search of hadoop against the snowflake domain and the term hadoop against the term snowflake, I keep finding that Snowflake has a definite Target in mind which is to convert hadoop users or people evaluating hadoop to choose them instead. They even have a webinar specifically for that segment of people.
Even further searching of Alex Woodie and mentions of snowflake show multiple articles with the CEO across multiple domains including datanami and Enterprise tech.
All that is circumstantial but I'm exercising a healthy bit of skepticism that this piece is pure research done by Alex Woodie. A little more objectively,
If I examine the "points" of the article, what I can see is:
Bob muglia has never met a happy hadoop customer. Mention couple of things that might replace hadoop in the future.
Bob muglia has only seen a few customers who've tamed hadoop.
Some discussions with and about Facebook's experience with hadoop painting hadoop as hard work from the outset.
More discussions with other tech folk (Kafka and data torrent). One is an alternative of sorts, and the other again discusses pain of hadoop.
And then back to Bob Muglia and who his target customers are for Snowflake - "hadoop refugees" - and his belief that we are in the valley of despair regarding hadoop.
Which brings us to the final mental point of the article. Ditch hadoop sooner rather than later, and here are the alternatives where the main one pushed from start to end is Snowflake.
I apologise if this was too far off the topic. I think the discussion of hadoop's validity or how it's being used is valid. I do also believe it's healthy to call out suspect stuff like this because the core of the article itself provides little to no critical value.
I guess that kind of makes sense. What programmers like might not necessarily be a good basis for the long-term. I do have to say, though, that part of the failure of Hadoop is it generated a lot of hype and so better tools and alternatives developed to meet that need. So when you're saying a better alternative is Spark or Kafka, I feel like it's almost as if you're saying "Oracle is a failure that never materialized the promised benefits. Instead we should be using PostGres and MySQL."
But I also know that with a lot of big data hype, businesses were wanting to do Hadoop and NoSQL and all this stuff because it was the cool new thing, not that they actually needed it. I've heard data scientists make the joke that every business thinks they need these tools because they're having difficulty running their business out of a spreadsheet.
I think it's important to remember that for most businesses, Spark, Hive, whatever; those aren't the right tools, either. SQL is still what most companies need. Businesses want machine learning, but usually what they need is boring old statistics. In an industry that always wants to be ahead of the curve, we tend to forget that it's not always the right thing to have the newest toys. Sometime companies do well utilizing the latest and greatest, but sometimes they just use what already exists wisely. I suspect that for many of those companies that used Hadoop and felt it didn't work them, the problem wasn't that they needed Spark instead, it was they were trying to solve problems that didn't exist. Man, I'm too young to sound this old. But, eh, yeah, we need to respect our elders' technology first, and consider the newest stuff only when we have a definable need for it.
They had heard Hadoop didn't require conforming to a schema and would allow data-driven insights. What they failed to verify in advance was whether those insights would increase profit.
I don't see how any other database-ish technology would change the situation. It ain't the tech, or the analysis, it's the business processes.
Example is Postgres. Now has with JSONB as document-oriented field type. Now has Postgres XL, as horizontally scalable ACID database. Will have, by approximately august', abiltity to maintain views in memory (aka lambda.architecture speed layer) via PipelineDB extension, for fast streaming analytics.
It seems that a combination of Postgres (with extensions) + Kafka +Redis -- is a strong stack for lambda architecture and, as initial data hub component of the overall puzzle
While Spark (or even Python+Dask) can be viewed as distributed data analytics platforms that replaces 'non-ui centric' BI. I think ui-centric BI (eg adhoc reports/visualizatiosn) are going to be still dominated within enterprises by Tableu/QlickView type of solutions.
For traditional BI-oriented data marts (that organized downstream from the datahub) -- probably traditional column oriented databases, and the new open source ones make sense.
To me ,the promise of hadoop being a silver bullet for 'all the big data needs' -- was always nothing more than unsubstantiated hype.
So it definitely failed the ones who believed, the hype -- but did not fail others who did not buy into it.
Hadoop (and its general ecosystem, which includes HBase), is a fairly good idea. Its core ideas - map/reduce on Hadoop, and a large distributed key/value store for HBase - are actually pretty solid.
And for many years, there were simply no alternatives to Hadoop. Think of the years 2008 to 2012/13. If you had to process terabytes or petabytes of data, what were your solutions? No wonder Yahoo and Facebook (and others) put in so much effort into their Hadoop solutions.
But, IMHO, there were several issues with Hadoop and their ilk.
1. The core infrastructure wasn't stable enough. Hadoop / HBase were supposed to be distributed systems, and they worked well, but small failures could bring down your entire cluster. Given that Hadoop and HBase were being used in mission-critical systems in the cloud, and given the amount of DevOps or sys-admin work that went into maintaining these, I'm not surprised people eventually migrated to distributed systems that were easier to maintain and run.
2. There are now plenty of "hosted on the cloud" solutions such as Amazon DynamoDB or similar cloud solutions. When your company depends on 99.99% or similar SLAs, you don't want to have downtime on your database systems and spend time debugging complicated core dumps on your Hadoop or HBase clusters when you can just store it "on the cloud" and be done with it. Sure, there's a higher price point, but those are the trade-offs you live with.
3. If you want to be in-house, there are plenty of alternatives out there as well today. Apache Spark for processing, Kafka for a messaging bus / streaming data, ElasticSearch for large scale storage, with multiple indices. Many of them are much more robust than Apache Hadoop / HBase, and I'm not surprised they've gotten more traction recently.
Ultimately, I think Hadoop / HBase are just showing their age. They were fantastic for the first wave of Big Data technologies, and you had little alternative if you were building large-scale systems circa 2008 to 2013, but now, you just have a plethora of choices from various vendors.
I think the comment "first wave of Big Data technologies" is interesting. I agree with it, but also, the term "big data" is relevant to a point in time, I think. I mean, VLDB was the term before "big data" - I think when the term "big data" became popular was when technology such as Hadoop became more mainstream and available instead of just being used in special cases.
I was working with terabytes of data in SQL Server prior to 2008 (around 2004/2005 - which was doable but pushing some limits) and then got into Hadoop etc. afterwards. It was interesting to reach the limits of a tech we were using and be there for the transition to Hadoop etc.
There were expensive commercial alternatives - MPP databases like Teradata, Netezza or Vertica - or things like scientific computing platforms (which were very limited to scientific computing scenarios). But nothing was as.. mainstream and available as Hadoop became. And yet, Hadoop never really completely beat the traditional MPP vendors until you started getting engines like Impala, Presto, Spark etc. that made SQL on top of Hadoop easier. And that took quite a while.. in fact I'd say making SQL on top of Hadoop work well (among many other things) is probably still on-going in a way. We are much, much further along with that now than we were.
On the other hand, it was interesting to see the new opportunities develop for working with data other than just SQL - straight MapReduce, machine learning, graph processing, DAG, etc.
Will be interesting to see where we go from here. I still think there is a lot going on.
To be honest, Hadoop on AWS through Elastic MapReduce (EMR) actually served us really well. The cluster downtime we had in our early days were all "human error" (configuration and Map/Reduce programming mistakes). Once we fixed those, EMR served us like a champ.
The bigger issues we had were with HBase actually, particularly during its region splitting. It was great for storing massive amounts of data that you could store and query fast (with the right key design), but we weren't able find people skilled enough in HBase ops to make us comfortable with continuing to use it.
I have seen Teradata, Netezza and Vertica being used heavily at that time (circa 2010-2013) but never worked on them directly. Like you said, they are extremely expensive, and I can imagine why some of the larger installations were trying to move to alternatives.
End of the day, I don't have any perfect solutions to recommend to Hadoop or HBase. I think there are a variety of cloud-based solutions from Amazon, Microsoft and Google; as well as hybrid open-source solutions that do the job, but to each his own!
I do agree hbase is a monster to manage. When it works though, there aren't many systems which can mach it's insert/scan capabilities T scale.
As for Kafka, I was surprised to see the article lump this together with Hadoop. I've used Kafka exclusively as a messaging bus or to stream large amounts of data. It doesn't exist as a replacement to Hadoop or Big Data technologies, but as a perfect complement, because it doesn't process data, just streams / moves it efficiently.
We can build better data infrastructure for data scientists, but in practice it's hard to sell "10x easier to use" into organizations with hadoop or half-broken bespoke infrastructure because the IT groups running the show don't really care that they're making their data scientists miserable.
If change is to come, it's going to have to be from data scientists embedded within business units demanding better tools, because hadoop works just fine if you don't care how hard it is for your users to access their data.
Also, those entities who the article claims have a grasp on Hadoop are some of the main people who helped develop it - lest we forget that the early work on Hadoop came out of Yahoo. "Web-scale" companies by their very nature have to own this technology, because almost nobody else on earth did what they have done on that scale (except maybe governments and scientists).
However, I agree the article is right on. As someone who has done a lot of work with different Hadoop distributions (Cloudera, Hortonworks and MapR to name a few), I can say that compared to the performance and ease of use of more traditional data platforms, Hadoop is a different and more complex creature. It has evolved quickly in some regards and slowly in others. It's way easier to get a Cloudera system up and running today (for example) than it was in the past.
"Hadoop" by itself is a misnomer, it's not like it's one technology. HDFS, MapReduce (and MR2), Hive, HBase, Spark, Pig, Flume, etc. etc. - it's an ecosystem. It's also a mashup of many things depending on what distro you're talking about - too many things - a "data hub", an "operational data store", a "pre-data warehouse staging", an ETL platform, a "search" platform, a data governance platform, a data archiving solution, a real-time data query / streaming data processing etc. etc. etc.
I think Hadoop did one important thing to our overall awareness, and that was to think of a "data platform", "data processing", "data architecture" and "data infrastructure" as a major, and "big" strategic part of an organization. I think the rise of "big data" thinking along with the rise of the "chief data officer", "data scientists" and "data engineers" and the like all goes together - a shift in thinking we made where we realized we could do more with more data (for better or worse).
There are many new and old technologies that help people implement the data infrastructure in their organizations - not just Hadoop. Hadoop just happened to be a major player at the epicenter at this (and open source as well), and if anything, has spurred technological advancement in databases, NoSQL and NewSQL and OldSQL alike, ETL, data mining, AI/machine learning, etc. So it has played its role, whether it specifically stays relevant or not as we perceive it.
The decline of Hadoop as a software category is Software Product Marketing 101: it did not identify pervasive killer use cases critical to running . Yes, it's true that Hadoop was a revolutionary way to store and process massive datasets on commodity hardware, but what's the use case for that? If you are Visa/AMEX (fraud detection), Facebook/Google (various ML-based data products) and a few other types of companies with obvious applications of massive data processing, yes, Hadoop has been great.
But here's the thing: beyond a few such corner cases, it never found a use case that enterprise data warehouse couldn't handle.
Then came Redshift, then BigQuery, and now Snowflake (as a BigQuery on AWS, really). While there are some key technical differences between Redshift and BigQuery/Snowflake, they are all _much_ cheaper than the previous generation of data warehouses (Vertica, Netezza, Greenplum, etc.) The lower price meant greater access, and developers who previously couldn't imagine using data warehouses could finally spin one up with a credit card swipe.
Hadoop, too, took a lot of collateral damage because many developers realized that they didn't need much of Hadoop beyond SQL-on-Hadoop.
Redshift was a beautiful feat of product strategy and marketing: They just took what used to cost a lot and offered it for much less in an environment where developers already had a lot of data (AWS). This was much simpler to execute than what Hadoop had to do: introduce new technology, identify use cases, and finally compete with incumbent solutions.
We failed Hadoop (as you can see from www.cloudera.com, even Cloudera, the Hadoop company, hardly mentions Hadoop on its top page). Not the other way around.
I wonder how many of those are just SQL-on-HDFS (Drill, Spark, Presto) ?
The trend I'm seeing the big Data sphere in my company is a by-and-large move away from technologies that implement MapReduce and and complicated batch processing with HDFS as a data store. More and more customers want insight as soon as they get/produce their data so we've seen a particularly large increase in interest in technologies such as Kafka/ Samza and spark/ pySpark.
I see a trend in Kafka and but I think the community needs to jump behind it too, keep it as a pipeline tool and not a querying engine.
I dont see Hadoop-based solutions going away any time soon though.