MidScale Solution Architecture
|MidScale solution architecture is a living document. It will be continuously updated during the course of the project.|
Goal of midScale project is to significantly improve performance and scalability of midPoint. We would like to make midPoint operate faster (performance). Even more importantly, we need midPoint to be scalable, i.e. it has to handle large amounts of data and requests without significant degradation of performance.
MidPoint is a popular and comprehensive open source identity management and governance platform. However, midPoint has one significant limitation. MidPoint was originally built to address the needs of mid-size enterprises, agencies and universities. Initial design of midPoint data store components favored flexibility and time to market. As midPoint was targeting mid-size organizations the scalability was not high on a list of implementation priorities. Now, midPoint is being deployed to handle scenarios with large number of identities. Deployments that manage students, subscribers and consumers are becoming more and more common. Which makes sense, as these types of users can especially benefit from the data protection capabilities of midPoint. However, such deployments are hitting scalability limitations of current data storage components of midPoint.
Future scalability issues were foreseen in original midPoint design. The datastore implementation (repository) was designed as replaceable component. Based on our previous experience, we have identifier current repository implementation as a major obstacle to scalability. There are other areas where we suspect performance and scalability limitations, such as Prism, our basic data representation layer. We expect to improve midPoint clustering, task management mechanisms and other areas.
MidScale is all about performance and scalability. As for the performance, our plan is to proceed systematically and premature optimizations. Therefore the initial phases of the project are focused on testing environment. We plan to create a testing environment and use it to identify performance and scalability problems. We will also use the testing environment and other analytic tools (e.g. profiling) to confirm performance issues that we are currently suspecting.
At the same time, we are conducting a series of design meetings to identify potential problems and outline solutions. Data from the testing and prototyping are fed back to the design process.
The testing environment will be used and even expanded during the entire duration of the project - and even permanently maintain after the project. The intermediate results from the testing will be used to improve the implementation, and even re-evaluate the design if needed.
Interative and IncrementalMidScale project is conducted in iterative and incremental manner. There are many iterations separated by milestones. The plan will be adapted as the project evolves in time. Therefore MidScale solution architecture is a living document. It will be updated during the course of the project to reflect our current thinking.
MidPoint is not bound to any particular data store or database. Thanks to such foresight, midPoint has a flexible and replaceable data storage components. We would like to take advantage of this design feature and re-implement data storage components in a scalable way.
The current database schema has many limitations, mostly caused by two reasons:
Current database schema is based on early design that prioritized functionality over scalability. MidPoint evolution emphasized database schema compatibility, which meant that sub-optimal early design is still limiting the perfomance and scalability of data model.
Database schema was designed in a generic way that had to fit several database engines (PostgreSQL, MySQL/MariaDB, Oracle, Microsoft SQL). Therefore the data model could not use any database-specific optimizations. The obvious method to support several databases at the time when midPoint was designed was to use object-relational mapping (ORM) component, such as Hibernate. This approach introduces additional limitations and difficulties.
Our plan for scalability improvement has several key aspects:
Focus on scalability and efficiency in database schema design. Design database tables and indexes in such a way that supports large data structures, efficient data storage ans searching.
Choose just one database to support. Preliminary choice is PostgreSQL, however, that choice has to be validated in early stages of the project. Choice of one database engine allows us to use database-specific scalability features and optimizations, e.g. hierarchical tables and partitioning. We can also better "tune" the code for high performance. Additionally, this choice will simplify testing and lower overall support cost, making the solution more maintainable in the long run.
Connect to database directly, avoiding ORM. The ORM layer (Hibernate) is not providing any significant advantage in this case. Moreover, it can be a source of problems, slow-downs and limitations.
Design of database schema is still work in progress. Initial tasks in the project are focus on evaluation of current problems, design of solutions and simple rapid prototypes.
Please see repository design notes for the details.
Even after midScale is finished, we still plan to keep the existing Hibernate-based repository implementation (referred to as "legacy" repository implementation). The plan is to use this implementation to support midPoint running on Oracle and Microsoft SQL databases (support for MySQL/MariaDB is already deprecated). MidPoint should be able to operate normally by using this repository implementation. The performance is likely to be low (approximately at current levels). Yet such performance is still sufficient for many (usually smaller) deployments. The plan is to support this "legacy" repository implementation at least for the lifetime of miPoint 4.4 LTS, which is planned to end in 2024. However, the future of the "legacy" implementation starting from midPoint 4.5 is yet to be decided.
MidPoint works with a high-level data model, which is currently defined by XML Schema (migration to Axiom planned in the future). Therefore a special query language is needed to form queries and search filters using this data model.
MidPoint has existing XML-based query language that originated in the early years of midPoint development. However, this XML method is at the end of its life for several reasons. Firstly, the language is XML-based, which is a major usability problem. As midPoint deployments grow larger and more complex, users need more complex methods to search, sort and report data. Users often need complex features that are difficult to provide in GUI. Therefore the users resort to the query language. However, the XML form of the language is a major usability obstacle. Secondly, as midPoint is turning away from XML, the language need to be transformed to JSON, YAML and possibly future data representation formats. While this is still possible, it adds additional complexity, while still not solving the usability issues. The query language is not a technological obstacle for scalability. However, the query language is a practical and psychological barrier for scalability. MidPoint could work well with the current query language even with massive data sets. However, human administrator will not be able to manage it.
Query language was a subject of many design discussions in the past. The preliminary result of such discussion was a desire to replace the query language with something that would be much more friendly to users. Major advancement in the design was made in mid-2020 when Axiom data modeling language was designed. The work on Axiom clearly indicated the direction of future development of midPoint data model, a direction that points away from XML. We see the future of data model in independence from data representation formats such as XML, JSON or YAML. This also indicated the direction for the query language.
The hard requirement is that the new query language must be Axiom-compatible. I.e. it must describe the same data structures, it must have the same general look and feel and it must the same concepts (e.g. item path). The new language must be reasonably user-friendly. The goal is not necessary to be friendly to common end users of the system. However, it has to be friendly, understandable and intuitive to identity administrators that are somehow familiar with midPoint data model.
The progress of the query language design is documented in query language design notes.
Horizontal scalability is a key to massive scaling. It cannot be expected that a single node can handle all the load, not to mention resilience and reliability.
MidPoint is built for horizontal scalability. However, there are significant improvements to be made in several areas.
MidPoint operations are conducted in a context of a task. MidPoint has a specialized task manager component, which is responsible for both intra-node and inter-node task management. Support for multi-threaded task execution was added to midPoint few years ago. However, there are still some issue and a room for improvement.
Multi-threading can help to utilize resources of a single midPoint node, however multi-node task distribution is needed to utilize resource of the entire cluster. MidPoint has a capability to distribute some tasks among the nodes of the cluster (known as "multi-node tasks"). However, this capability is somehow static, difficult to configure and error-prone. The plan is to make the tasks adapt to cluster conditions automatically or semi-automatically. The specific mechanism is not clear yet, it will be designed and prototyped in a later part of the project, during an "autoscaling" activities. The autoscaling will also address the ability to dynamically add and remove nodes to midPoint cluster, improving on a current static cluster configuration.
Most of the horizontal scalability improvements are likely to impact the task manager component of midPoint. Task manager is responsible for scheduling, running and controlling execution of midPoint tasks. This includes the ability to distribute tasks in a cluster. Therefore the task manager component is the key to horizontal scalability.
Prism component forms an underlying foundation under midPoint data model. Prism data structures are used by all midPoint components, and it is used often. Therefore, even a small inefficiency in Prism code can multiply and cause serious issues. We suspect several issues in the Prism code already. However, these need to be measured and confirmed, as we want to avoid pre-mature optimizations.
Moreover, we suspect stability issues in Prism code, especially related to thread safety. We have observed instability of Prism code under heavy multi-threaded use. This is a major obstacle to scalability, as heavy multi-threaded use of midPoint increases the probability of failures and errors.
Prism design has few gaps when it comes to thread safety, which needs to be improved. The Prism design discussions uncovered several potential problems and suggested solutions. The design work will continue as the implementation and testing of midScale projects progresses.
There is also concern of Prism API stability. Prism is used in midPoint code, but it is also used by midPoint extensions. Complex, large-scale deployments are likely to need special-purpose extensions. Prism API is not yet "finalized", it is still evolving and changing. While the changes tend to be small, there is yet no compatibility guarantee for extensions. Stabilization of Prism API is expected to be one of the side effects of midPrivacy efforts, which will be an indirect benefit for complex midPoint deployments.
Please see Prism Design Notes for the details.
There are other aspects to scalability besides raw computing and data processing power. The system has to deal with a massive amount of data and present the data to user. Therefore midPoint has to be able to present large data sets in an efficient and user-friendly way.
MidPoint has complex administration user interface built on Apache Wicket framework. As midPoint was originally designed to handle smaller number of identities, GUI performance was not a priority. There are several known and many unknown performance issues in current GUI implementation, many of them originate from an incorrect use of the Wicket framework. The plan is to identify these issues using a combination of measurements and code reviews. Once identified, we will design the way how to resolve the issues. However, most issues should be resolved by correcting the code to follow proper Wicket procedures.
When it comes to the user interface, there is also another aspect of scalability: user experience (UX). User experience design of midPoint user interface have not focused on the characteristics of massive data processing. MidPoint would greatly benefit from user experience improvements that can make administration of millions of identities easier. The best way would be to conduct a systemic assessment of midPoint user experience and design an overall UX approach. However, midPoint user interface is complex and such a task would not fit into the scope of midScale project. Yet, we still need to make some user experience improvements to make midPoint usable for large identity populations. We plan to use feedback from midPoint users and testers as a basis for designing the necessary UX improvements.
Please see GUI Design Notes for the details.
What is measured, improves. Therefore there is one area of midScale project completely dedicated to testing and measurements.
Most of the testing and measurement effort is focused on performance testing environment. This is a dedicated environment for midPoint performance and scalability testing, which allows to set up various test cases involving large data sets and user load. This will be a primary tool to assess progress and results of the project.
The performance testing environment is not the only testing tool that we will employ during midScale project. There is an existing testing suite of midPoint tests, which is composed mostly of integration tests. We plan to utilize this suite, at least to get some informational data.
There is also an existing GUI testing framework code-named "Schrodinger". The Schrodinger framework will need improvements in order to work for midScale project. We plan to make the necessary improvements and add Schrodinger to the set of midScale testing tools.
It is essential the steps taken in midScale project are aligned with a long-term vision of midPoint development. Therefore we have conducted a related design activity to improve vision for midPoint 5. MidPoint 5 is a planned major midPoint release, which will be start of a next generation in midPoint development.
Please see MidPoint 5 Vision for the details.