Tez generalizes the MapReduce paradigm to a more powerful framework based on expressing computations as a dataflow graph. Tez is not meant directly for end-users – in fact it enables developers to build end-user applications with much better performance and flexibility. Hadoop has traditionally been a batch-processing platform for large amounts of data. However, there are a lot of use cases for near-real-time performance of query processing. There are also several workloads, such as Machine Learning, which do not fit will into the MapReduce paradigm. Tez helps Hadoop address these use cases.
The main reason for Tez to exist is to get around limitations imposed by MapReduce. Other than being limited to writing mappers and reducers, there are other inefficiencies in force-fitting all kinds of computations into this paradigm – for e.g. HDFS is used to store temporary data between multiple MR jobs, which is an overhead. (In Hive, this is common when queries require multiple shuffles on keys without correlation, such as with join - grp by - window function - order by.)
The key elements forming the design philosophy behind Tez -
Empowering developers (and hence end users) to do what they want in the most efficient manner
Expressive Dataflow APIs - The Tez team wants to have an expressive-dataflow-definition API so that you can describe the Direct Acyclic Graph (DAG) of computation that you want to run. For this, Tez has a structural kind of API in which you add all processors and edges and visualize what you are actually constructing.
Flexible Input-Processor-Output runtime model – can construct runtime executors dynamically by connecting different inputs, processors and outputs.
Data type agnostic – only concerned with movement of data, not with the data format (key-value pairs, tuple oriented formats, etc)
Simple Deployment – Tez is completely a client-side application, leverages YARN local resources and distributed cache. There's no need to deploy anything on your cluster as far as using Tez is concerned. You just upload the relevant Tez libraries to HDFS then use your Tez client to submit with those libraries.
You can even have two copies of the libraries on your cluster. One would be a production copy, which is the stable version and which all your production jobs use. Your users can experiment with a second copy, the latest version of Tez. And they will not interfere with each other.
Tez can run any MR job without any modification. This allows for stage-wise migration of tools that currently depend on MR.
Exploring the Expressive Dataflow APIs in detail - what can you do with this? For e.g. instead of using multiple MapReduce jobs, you can use the MRR pattern, such that a single map has multiple reduce stages; this can allow streaming of data from one processor to another to another, without writing anything to HDFS (it will be written to disk only for check-pointing), leading to much better performance. The below diagrams demonstrate this -
The first diagram demonstrates a process that has multiple MR jobs, each storing intermediate results to the HDFS – the reducers of the previous step feeding the mappers of the next step. The second diagram shows how with Tez, the same processing can be done in just one job, with no need to access HDFS in between.
Tez’s flexibility means that it requires a bit more effort than MapReduce to start consuming; there's a bit more API and a bit more processing logic that you need to implement. This is fine since it is not an end-user application like MapReduce; it is designed to let developers build end-user applications on top of it.
Given that overview of Tez and its broad goals, let's try to understand the actual APIs.
The Tez API has the following components –
DAG (Directed Acyclic Graph) – defines the overall job. One DAG object corresponds to one job
Vertex – defines the user logic along with the resources and the environment needed to execute the user logic. One Vertex corresponds to one step in the job
Edge – defines the connection between producer and consumer vertices.
Edges need to be assigned properties; these properties are essential for Tez to be able to expand that logical graph at runtime into the physical set of tasks that can be done in parallel on the cluster. There are several such properties –
The data-movement property defines how data moves from a producer to a consumer.
Scheduling properties (sequential or concurrent) helps us define when the producer and consumer tasks can be scheduled relative to each other.
Data-source property (persisted, reliable or ephemeral), defines the lifetime or durability of the output produced by our task so that we can determine when we can terminate it.
You can viewthis Hortonworks articleto see an example of the API in action, more detail about these properties and how the logical graph expands at run-time.
Theruntime API is based on an input-processor-output modelwhich allows all inputs and outputs to be pluggable. To facilitate this, Tez uses an event-based model in order to communicate between tasks and the system, and between various components. Events are used to pass information such as task failures to the required components, flow of data from Output to the Input such as location of data that it generates, enabling run-time changes to the DAG execution plan, etc.
Tez also comes with various Input and Output processors out-of-the-box.
The expressive API allows higher language (such as Hive) writers to elegantly transform their queries into Tez jobs.
The Tez scheduler considers a lot of things when deciding on task assignments – task-locality requirements, compatibility of containers, total available resources on the cluster, priority of pending task requests, automatic parallelization, freeing up resources that the application cannot use anymore (because the data is not local to it) etc. It also maintains a connection pool of pre-warmed JVMs with shared registry objects. The application can choose to store different kinds of pre-computed information in those shared registry objects so that they can be reused without having to recompute them later on, and this shared set of connections and container-pool resources can run those tasks very fast.
Overall, Tez provides a great deal of flexibility for developers to deal with complex processing logic. This can be illustrated with one example of how Hive is able to leverage Tez.
Let's take this typical TPC-DS query pattern in which you are joining multiple tables with a fact table. Most optimizers and query systems can do what is there in the top-right corner: if the dimension tables are small, then they can broadcast-join all of them with the large fact table, and you can do that same thing on Tez.
But what if these broadcasts have user-defined functions that are expensive to compute? You may not be able to do all of that this way. You may have to break up your tasks into different stages, and that's what the left-side topology shows you. The first dimension table is broadcast-joined with the fact table. The result is then broadcast-joined with the second dimension table.
Here, the third dimension table is not broadcastable because it is too large. You can choose to do a shuffle join, and Tez can efficiently navigate the topology without falling over just because you can't do the top-right one.
The two benefits for this kind of Hive query with Tez are:
it gives you full DAG support and does a lot automatically on the cluster so that it can fully utilize the parallelism that is available in the cluster; as already discussed above, this means there is no need for reading/writing from HDFS between multiple MR jobs, all the computation can be done in a single Tez job.
This particular Hive query is seeing performance improvement of more than 100% with the new Tez engine.
Richer DAG support. For example, can Samza use Tez as a substrate on which to build the application? It needs some support in order for Tez to handle Samza’s core scheduling and streaming requirements. The Tez team wants to explore how we would enable those kinds of connection patterns in our DAGs. They also want more fault-tolerance support, more efficient data transfer for further performance optimization, and improved session performance.
Given that these DAGs can get arbitrarily complex, we need a lot of automatic tooling to help the users understand their performance bottlenecks
Tez is a distributed execution framework that works on computations represented as dataflow graphs. It maps naturally to higher-level declarative languages like Hive, Pig, Cascading, etc. It's designed to have highly customizable execution architecture so that we can make dynamic performance optimizations at runtime based on real information about the data and the resources. The framework itself automatically determines a lot of the hard stuff, allowing it to work right out-of-the-box.
You get good performance and efficiency out-of-the-box. Tez aims to address the broad spectrum of use cases in the data-processing domain in Hadoop, ranging from latency to complexity of the execution. It is an open-source project. Tez works, Saha and Murthy suggest, and is already being used by Hive and Pig.
About the Authors
Arun Murthyis the lead of the MapReduce project in Apache Hadoop where he has been a full-time contributor to Apache Hadoop since its inception in 2006. He is a long-time committer and member of the Apache Hadoop PMC and jointly holds the current world sorting record using Apache Hadoop. Prior to co-founding Hortonworks, Arun was responsible for all MapReduce code and configuration deployed across the 42,000+ servers at Yahoo!
Bikas Sahahas been working on Apache Hadoop for over a year and is a committer on the project. He has been a key contributor in making Hadoop run natively on Windows and has focused on YARN and the Hadoop compute stack. Prior to Hadoop, he has worked extensively on the Dryad distributed data processing framework that runs on some of the worlds largest clusters as part of Microsoft Bing infrastructure