Best Practices for Designing Models: Difference between revisions

From QPR ProcessAnalyzer Wiki
Jump to navigation Jump to search
 
(25 intermediate revisions by the same user not shown)
Line 1: Line 1:
== Best Practices for designing dashboards ==
This page describes common best practices for designing a suitable structure for a process mining model, and how to configure the model settings. Best practices how to write ETL scripts that actually create and update models, are described separately.


=== Visualization and usability best practices ===
== Datatable column data types ==
* Use conditional formattings to improve KPI visualization
* Always use the most suitable [[Importing Data to Datatable from CSV File#Data types|datatypes]] for datatable columns, as the datatypes have remarkable performance impacts and they also affects how data can be used in the analysis. Datatable column datatypes will also be the case and event attribute datatypes in the model. As a general rule, avoid the ''string'' datatype when other datatypes can be be used. Here are some guidelines:
* Use on-screen settings for settings that often users want to change, as they are easier to use than opening the settings. They also guide users to change parameters that might be relevant from the analysis viewpoint.
** If there are only two possible values, ''boolean'' is the best datatype. The values in boolean are called ''true'' and ''false'' which can be mapped into a textual presentation in charts. Thus, it's not needed to use string datatype to get desired textual presentations in dashboards.
* Check that each measure and dimension has describing units. The general terms "cases" and "events" might not describe the counts best. E.g. cases might be orders.
** If there is numerical data that doesn't contain decimals or precision with decimals is not required, ''integer'' is the best datatype.
* Use custom labels if they describe the measures and dimensions better. Still, in many measures and dimensions, the automatically generated title is suitable.
** For timestamps, the string datatype will definitely not work, so make sure to use ''date'' type and the conversion from a textual value during the import interprets the data correctly. Even though it's not about a precise timestamp, but the precision is for example a day, the date datatype is still be best.
* Note the special values, such as null and empty strings, and set a describing label name for them. E.g. ...
** If the data contains a numerical score (such as number between 1 and 5), integer is better than string.
* Disable creating filters from chart, if there are no meaningful filters created.
* All datatypes support ''null'' values to mark missing or not existing value. The null value can be used to mark anything – its meaning is just a matter of decision. For not existing numerical values, using null is better than zero, as nulls are ignored in calculations (such as in average). Note that strings can also contain the empty string value, which is different than the null value. In addition, booleans can actually contain three values: true, false and null.
* Limit the number of shown attributes or event types, if there are some that are not needed. This doesn't have performance impact, though.


=== Performance optimization best practices ===
== Performance optimization ==
* As Analyzed objects, prefer Cases over Events, as usually there are lot more events than cases. Some KPI's can be calculated from the cases point of view. Also Variations, Event types and Flows are generally fast. On the other hand, Flow Occurrences is slow, as the number of them is even more than the event count.  
* Include to the model only those case and event attributes that are used in the dashboards, because loading the model is slower and it uses more memory, when there are more attributes. For advanced analysis, such as finding root causes and clustering, more attributes maybe useful, but not for dashboards using only the specified attributes. The calculation performance itself doesn't deteriorate, even though the number of attributes increase.
* Prefer ready-made measures and dimensions over custom. For some simple calculations, Statistical calculations may be used. Also the Adjustment expression is useful.
* Include only those event types to models, that are needed by the dashboards and analyses. The more there are events, the more model loading takes, the more model uses memory, and the more calculations take. Event types can be excluded using filters, improving the measure calculation performance similar to corresponding smaller model, but on the other hand calculating the the filter take time.
* Limit number of returned rows: The Max rows setting determines the number of returned rows. The less there are rows the better performance. Usually in dashboards, limiting the amount of shown data is desired also for usability, e.g., show only the top-20 items etc.
* For large models, the [[Automatic_Model_Loading_on_Server_Startup|Load Model on Startup]] setting may be useful, so that the initial opening of a dashboard doesn't take too long, as with the setting the model is already available in the memory. On the other hand, pre-loading many models automatically, consume more memory, so models that are not used regularly, should not be loaded automatically into memory.
* Group rows exceeding maximum affects performance, so use it only when the information is useful for the analysis.
* When new data is loaded into the model, it will become progressively slower to use. If the oldest data in the model is not shown by dashboards, it's recommended to remove it from the model at the same time when new data is loaded. The old data may be stored to an archive model, if it's still needed in some situations.
* Sorting affects performance, so use it only when it's relevant for the analysis
* Note the number of charts in a dashboard. The more there are charts, the slower it is.
* Alternative to chart filter is Analyzed objects containing filtering, e.g. ... might improve performance
* Don't use dimensioning when it's not needed. When there is anyways a row for each root object, dimensioning is unnecessary. For example, Cases as Analyzed objects and dimension by case id will lead to a row for each case, but the same result can be achieved by disabling dimensioning.


=== Advanced performance optimization best practices ===
== Usability for dashboarding ==
* For slow charts, use the Benchmark Performance to find the fastest settings. Usually settings up a working chart is the first thing to do, and if the chart appears too slow, you can try to find another, faster way to calculate the same chart.
* Use concise names for event types, as shorter names are easier to read by users and they also provide slightly better performance. This is also valid for case and event attributes values. For the readability, it's also better that if there are names that are close to each other, the differences would be in the beginning of the name rather than in the end, as the ending may be cropped out if there is lack of screen space.
* Avoid calculating same things multiple times in different measures. If there are repeating expressions, create a separate measure for it, and define it as a variable, which can be referenced from other measures.
* Use the model description to document any relevant details regarding the model for other users. For example, meaning of the event types and case/event attributes is definitely not obvious for other users but also perhaps not for the author itself after some time. The model description field can be found in the [[QPR_ProcessAnalyzer_Project_Workspace#Editing_Model_Settings|Model Properties dialog]].
* Same dashboard can easily use different models and filtering still works. Model optimized for a chart might improve performance.
* When data is sorted their data types are important, for example numerical data is sorted by the number values, where as strings are sorted alphabetically. The difference is, for example when sorting numbers 9 and 10 ascending, the 9 is before 10, but if data is stored as strings, the "10" is before "9". If strings need to be in specific order, this needs to be taken into account when formatting data. For example the string values can be prefixed with an order number. The previous example with 9 and 10 can be worked around by storing string values "09" and "10".
* Try sampling. It improves performance, but in most cases, it cannot be used, as it affects the analysis results, for example in object counts. When sampling can be used, it's very useful in improving calculation performance for large models.
 
=== Other best practices ===
* '''Use preset as basis''': The preset contain commonly needed analysis, so in many cases, you find what you are looking for from the presets. It's easy to take a preset as a basis, and continue modifying the chart settings for your customized needs.
* Mappings can be done freely, so dimensions don't always need to go to the X-axis and measures go to the Y-axis.
* Avoid Custom layout settings as their compatibility with future QPR ProcessAnalyzer versions might not be maintained. Use Custom layout only when it's absolutely necessary for the visualization.
* '''For large data exports use CSV''': When there is need to export large amount of data, prefer the CSV export over Excel export, because the CSV performs better for large data.
 
== Best practices for creating models ==
* Use the most suitable datatypes for case and event attributes. If there are only two possible values, boolean is the best. The true and false values can be mapped into a textual presentation, so it's not needed to use strings to get desired texts for visualizations. If numerical data cannot contain decimals or precision containing decimals is not required for the analysis, integer should be used over float. If the attribute value contains a numerical score (such as number between 1 and 5), integer is better than string. Usually string is the slowest.
* All datatypes support null values to mark missing or some kind of special values. The null value can be freely used to mark anything - it's just a matter of decision.
* Include only case and event attributes that are needed by the dashboards. For analysis, more attributes maybe useful, but they are not needed for dashboards. Loading model is slower, when there are more attributes.
* Note the Load Model on Startup setting. When to use it correctly.
* Include only events that are needed by the dashboards
* Shorter event type names are easier to read in the UI and provide slightly better performance. This is also true for case and event attributes values.
* Use calculated attributes, to pre-calculate case level KPI's from measures. It cannot be used when there is event type filtering applied. On the other hand, don't use calculated attributes unnecessarily because they are stored into memory, and thus they consume memory list the normal attributes. Don't calculate anything from the entire model level in the calculated attributes expression, because it will lead to very slow performance in model loading.
* Use the model description to document the necessary details regarding the model for other users.
 
== Best practices for writing ETL scripts ==
 
== Best Practices for hosting QPR ProcessAnalyzer ==

Latest revision as of 09:35, 31 March 2022

This page describes common best practices for designing a suitable structure for a process mining model, and how to configure the model settings. Best practices how to write ETL scripts that actually create and update models, are described separately.

Datatable column data types

  • Always use the most suitable datatypes for datatable columns, as the datatypes have remarkable performance impacts and they also affects how data can be used in the analysis. Datatable column datatypes will also be the case and event attribute datatypes in the model. As a general rule, avoid the string datatype when other datatypes can be be used. Here are some guidelines:
    • If there are only two possible values, boolean is the best datatype. The values in boolean are called true and false which can be mapped into a textual presentation in charts. Thus, it's not needed to use string datatype to get desired textual presentations in dashboards.
    • If there is numerical data that doesn't contain decimals or precision with decimals is not required, integer is the best datatype.
    • For timestamps, the string datatype will definitely not work, so make sure to use date type and the conversion from a textual value during the import interprets the data correctly. Even though it's not about a precise timestamp, but the precision is for example a day, the date datatype is still be best.
    • If the data contains a numerical score (such as number between 1 and 5), integer is better than string.
  • All datatypes support null values to mark missing or not existing value. The null value can be used to mark anything – its meaning is just a matter of decision. For not existing numerical values, using null is better than zero, as nulls are ignored in calculations (such as in average). Note that strings can also contain the empty string value, which is different than the null value. In addition, booleans can actually contain three values: true, false and null.

Performance optimization

  • Include to the model only those case and event attributes that are used in the dashboards, because loading the model is slower and it uses more memory, when there are more attributes. For advanced analysis, such as finding root causes and clustering, more attributes maybe useful, but not for dashboards using only the specified attributes. The calculation performance itself doesn't deteriorate, even though the number of attributes increase.
  • Include only those event types to models, that are needed by the dashboards and analyses. The more there are events, the more model loading takes, the more model uses memory, and the more calculations take. Event types can be excluded using filters, improving the measure calculation performance similar to corresponding smaller model, but on the other hand calculating the the filter take time.
  • For large models, the Load Model on Startup setting may be useful, so that the initial opening of a dashboard doesn't take too long, as with the setting the model is already available in the memory. On the other hand, pre-loading many models automatically, consume more memory, so models that are not used regularly, should not be loaded automatically into memory.
  • When new data is loaded into the model, it will become progressively slower to use. If the oldest data in the model is not shown by dashboards, it's recommended to remove it from the model at the same time when new data is loaded. The old data may be stored to an archive model, if it's still needed in some situations.

Usability for dashboarding

  • Use concise names for event types, as shorter names are easier to read by users and they also provide slightly better performance. This is also valid for case and event attributes values. For the readability, it's also better that if there are names that are close to each other, the differences would be in the beginning of the name rather than in the end, as the ending may be cropped out if there is lack of screen space.
  • Use the model description to document any relevant details regarding the model for other users. For example, meaning of the event types and case/event attributes is definitely not obvious for other users but also perhaps not for the author itself after some time. The model description field can be found in the Model Properties dialog.
  • When data is sorted their data types are important, for example numerical data is sorted by the number values, where as strings are sorted alphabetically. The difference is, for example when sorting numbers 9 and 10 ascending, the 9 is before 10, but if data is stored as strings, the "10" is before "9". If strings need to be in specific order, this needs to be taken into account when formatting data. For example the string values can be prefixed with an order number. The previous example with 9 and 10 can be worked around by storing string values "09" and "10".