On the page in Microsoft docs where they discuss API Service Protections there is towards the end of the page a part which gives some recommendations. Some are great, like the recommendation to use many threads and remove the affinity cookie, however when I read it I really bounced at the recommendation that batching shouldn’t be used. That just didn’t rime with my experiences of doing heavy dataloads to dataverse. So I thought I might just test to see if it was true or not by creating a simple script in SSIS with Kingswaysoft. My results, using batching compared to not using it gives more than a 10x performance increase. Continue reading to understand more about how I tested this and some deeper analysis.
Parameters and excel
The first thing I did was to create an excel sheet for storing all the results. I really did have to think about the different parameters that could affect the result, so I chose the following columns:
Dataload – how many records. This needed to be a bit larger to make sure that the throttling time of 5 min was passed.
Operation – Different dataverse operation take different amounts of time. For instance, creates are typically rather fast, but deletes, depending on table, can be a lot slower as the platform might execute cascading deletes based on one single delete. For instance, if you remove a contact with 100 tasks connected to it with the regarding relation set to “parental” or “cascade delete” it will actually remove all the 100 tasks. If set to “remove link”, the platform has to make an update to each of the tasks, removing the link. There are also special operations like merge which are rather complex.
Table – There is a large difference between the different tables. Some of the OOB tables have a lot of built in logic and really small non-activity custom tables can be a lot quicker to create, update or delete.
Threads – How many threads were used.
Batch – The size of the batches being used.
Duration / Duration (ms) – Duration is where I input the duration as a normal time. I created a calculation to calculate the corresponding amount of milliseconds.
Time per record (ms) – This is the division of the duration in ms with the total number of records. During this first test, I always used 100 000 records as the dataload, but it could be interesting in the future to see the differences between different dataloads, with all else being the same. This is also the main output from this test.
Strategy – It is possible to have different strategies. In this first version I just ran everything at once, hence I called the strategy “All at once”. Different strategies might be “5 on, 5 off”, meaning that you design the script to run superfast for 5 minutes, the throttling limit, and then stop and do nothing for 5 minutes and then loop this. Not always possible to use that kind of strategy, but for massive deletes of for instance market list members (cannot be removed with bulk delete) that might be an option.
API – There are currently two APIs that can be used. The new WebAPI which uses JSON payloads and the older SOAP API which used XML payloads. It stands to reason that the smaller JSON payload should cause the WebAPI to be faster than the corresponding SOAP API. However SSL encryption also causes the data to be compressed, which might make these differences smaller than expected. There is also a server side aspect to this, as the APIs will run through different parts of the code on the server side which could affect the performance.
No of columns – How many columns are being sent to the API. Of course there would be a difference if you send a create message with 3 columns compared to 30. Hence this is a relevant point. It is still a bit rough, as there is a huge difference in creating a boolean record, a 2000 character nvarchar or a lookup. This could also be something that was adapted.
Existing records – How many records existed in the system prior to running this? Not sure if this makes any difference, in other words, everything else equal, would it take more time to write 100k records to a system with 0 records or one with 10M records? As I don’t know, and cannot rule it out, I added it.
Latency (ms) – Daniel Cai, Founder of Kingswaysoft, always recommend that the SSIS script with Kingswaysoft be run “as close as possible to the dataverse”. That does in other words imply that the latency to the server affect the performance. Do calculate this, I used diag.aspx from the computer running the script.
Location – Which geo is the instance located in. This is more for general information, the latency is really the important factor here. The throughput might also have some affect if you are using a really bad line to the dataverse. I was using a wired 1 GBit line. In this test, I was using an instance I got hold of as MVP, which is located in the US and my own stationary computer at home (a AMD Ryzen 9 3900X 12-Core Processor 3.79 GHz with 32 GB of memory). Hence the latency was rather high and not in line with Daniel Cai’s recommendations. It is hence also something to investigate further.
No of users – As I, and some others in the community have described, throttling is based on a per-user and per-front end server basis. Hence utilizing several service principals/application users can effectivly multiply the throughput. In this test I used just one.
Instance type – It is well known that sandbox instances do not have the same performance as a production instance. If you find Microsoft support on a happy day and you are working with a larger (no of licenses) instance, you might also get them to relax the throttles a bit, especially if you mention that you are doing a migration. As these factors strongly affect the performance of large dataloads, I did have to add this. During this test I was using a non-enhanced production instance, in other words, a production instance on which no throttles had been relaxed.
DB Version – The final parameter that I thought might affect this is the actual version of the dataverse instance. As improvements and god forbidd sub optimal “improvements”, can cause enhancements or degradations of the performance, this is necessary to document.
SSIS/Kingswaysoft setup
For setup of create tests in SSIS with the Kingswaysoft addons I used a dataspawner (productivity pack) to generate the data. I then just sent this directly to the CDS Destination.
And the Data Spawner config:
And the CDS Destination config
The main changes done in this case were to the parameters “Batch size” (set to 20 in the picture above) and how many threads to use (also set to 20 in the picture above).
After each run, I checked the log from SSIS to see how long the entire process took. Due to the fact that I have a computer with many threads and for this case, enough memory, it is my perception that most of the threads allocated were also used.
Results
What are the results? This is a picture of the excel:
As you can see I did try both Create and Delete operations, but the results are rather obvious.
20 threads/20 per batch of both create and delete, took around 45 minutes
Reducing to 16/10 made only a minor difference – 48 minutes
Microsofts recommendation of not using batching, ie 20 threads/1 batch – took over 10 h, both for delete and create.
Using only 1 thread with 1 batch was more or less the same as using 20/1
1 thread with 20 in every batch (1/20) took almost 5 h, which is around half the 1/1 or 20/1
I think the results clearly show, that Microsoft docs are currently incorrect in their recommendation to not use batching. Perhaps they will update this soon. From an entitlement perspective, one needs to understand the additional cost of the “batch unpacking” request that is made. With 20 in every batch, this is an overhead of 1/21 but if you would lower the batches to 4, it would be 1/5. Hence using as large batch as you can, without loosing performance, is generally what I would recommend.
As I have implied in this article, there are a lot of other parameters to investigate in the API. I have a hunch that a create with 10 lookups compared to 10 textfields, will also make a significant difference, but I will need to test it.
Also do consider the request timeout. When working with complex and large batches, one request may taker quite some time. You will know, however, as it will return a timeout exception if you exceed it. Note that some records in that batch may have been written anyway. Just that your client wasn’t waiting around for the answer.
I do also encourage others to try out other parameters in the API. What is really optimal from many different aspects. From a mathematical perspective this can really be seen as a multidimentional surface where we are attempting to find the highest points. I have now started this journey, and I hope it was an interesting read. Please leave a comment, if you have any experience to share or just want to comment.
I was recently in charge of a large migration. It all went fine but not without hickups that typically are connected to moving large amounts of data to dataverse. We were using SSIS with Kingswaysoft and ended up using a local SQL database as staging database too. This article will discuss the different lessons learned and give some concrete tips when doing similar migrations.
One of my more popular articles is the article that describes how to optimize the writing of data to Dataverse/CDS. If you are working with migration of large amounts of data, as I will be describing here, I do suggest you have a look at it: https://powerplatform.se/fast-data-management-in-a-limited-cds-world/. I will not discuss those concepts in any detail here but we did use all aspects mentioned in that article.
I recently was in charge of a migration which used CSV-file exports from an old German system (with German field names!) which had many millions of records, in both large tables like “Contact” and “Sales Order”. However, the system we migrated from had a completely different data modell than the one used in Dynamics. For instance, each row describing a “Flight” had to be divided into two rows, one for the outgoing flight and one for the homecoming flight, in the order detail table. We also had to create a lot of related data which was referenced from the “Flight” table, for example location, agent and brand. In other words, there was quite a lot of heavy transformations going on and a lot of logic involved, such as change format on the old data to match the Dataverse model and apply rules to resolve old issues, such as bugs.
Initially we only got a quite small subset of the entire database load, and we started our migration journey by creating all the migration logic in SSIS (which facilitates the script and makes updates easy to handle). The script did include some functions that “joined” rather large tables, both from the CSV files but also related data fetched from the Dataverse based on primary and alternate keys. I was clear with the customer from the very beginning that I wanted a full export with the same amount of data that we could expect in the final migration, mainly for the opportunity to stress-test the SSIS script before the migration to the production environment took place and after a while we got the big files…
…And this was when the excrement hit the wind generator. The afore mentioned lookups just stalled forever. We noted that having a lookup (using Kingswaysoft Premium Lookup) works fine on a computer with 16 GB memory up to a few 100k of records. However, once the data starts reaching 500k and more, it just stalls forever (and don’t even get me start on the sort tools…). Not sure exactly if it would have been possible to fix this by adding more cores and memory, we didn’t try. We hence had to rewrite the script and implement a staging database instead. What we found, is that a dataflow with 1M+ records of lookups will be 100x faster if you import the data into SQL and do a join instead. Lookups still works for smaller tables and I am not against them per se, as they do make the migration simpler. Adding more tables to a migration database will increase complexity, and if you want to add a column in a table, that table do not only has to be added to one SSIS dataflow, but probably a few more. And you also must do an ALTER TABLE in SQL to add the field there too. It is therefore important to have a good mapping set before you start to create the script. And keep the complexity as simple as possible. You can also use SQL tasks in the migration script to update the tables straight after you read them to the staging database per automation, if you need to apply some kind of rules after the read to the staging database, and find it easiest with an SQL query.
The method we used for developing the migration was to first make a “skeleton” migration, based on the target data model. In other words, we started with trying to get a few of the easiest fields, not all, from all tables that was to be involved in the migration – maybe it could be called – model-first-approach, instead of starting with one table, completing this and then moving on to the next. The advantage of the model-first-approach is that you quite early can start some tests on the data, for instance setting up some quantitative test by checking in the source system for the quantities of contacts and then comparing these quantities to the target. The tests can typically be done by other people than the people building the migration scripts and hence this methods scales a lot better than table-by-table-approach. It is also possible for several devs to work in parallell with different tasks. Typically the more senior will build the skeleton and then more junior can add fields by field to each respective table. A negative aspect of this approach is that it requires a lot of re-loads (keep in mind that this was a first migration, so there are no prior data in the Dataverse that we needed to consider) and re-mapping. And it may be easier to “fall out of” the structure, if you just need “to add a little bit here and there”. It is however indeed hard to go table for table, especially with related data. If you already have a lot of live data, you should think about a way to easy identify the migrated data so you can bulk-deleted. And do not forget to engage the client early with raised questions and the mapping to make sure you have understood everything correctly and avoid unnecessary errors.
We also tried to create unique row identities that strictly was based on the source data. This is very useful as that allows for delta-migration, or to continue where we left off in case of a problem. Let’s say for instance that you want to migrate 3 million contacts. If, after 2.1 Million contacts the script breaks for some reason, it is good to be able to continue at 2.1M instead of restarting. In this case we didn’t use modifiedon-date to be able to do a full delta migration logic but it is certainly possible. For this we used the cache-transforms, easily fetch the already migrated data (if any) with the unique and sort out the already migrated data if it matched the key.
Another pattern that we used was that, after creating a specific record, like contact, we reimported the recordid (in this case contactid) together with the legacyid. This allowed us to directly join with this table when later adding tables with dependencies like lookups towards the contact table, could be joined with this mapping table so that we directly got the contactid when querying the related table.
Tips
When migrating from CSV, import them directly as source tables in the staging database. That way, in case you need to fix something, you have a good reference for quantities.
Get an example of the full data load as early as possible. A script that works for a subset might not work at all for the full dataload as was the case for us.
Automate as much as possible. Don’t use any hardcoded values that are environent specific, such as transactioncurrencyid, but rather read these to small tables or to SSIS variables. Use SQL Truncate to remove all data quickly in a table, and make this part of the SSIS script as an SQL task at the appropriate stage.
Always check the quantities. How many rows in source data, how many rows after a match and check if it differs so you very early can identify bugs in your script that might be the reason for dropping rows. For example, you might use a JOIN when you should use an OUTER JOIN. Always check the total number and see if it is what you expect. Watch out for duplicates, and always check so your unique IDs (if you got some from the source data) really are unique and not NULL. Do note that if you have duplicates, that you join on, that will create multiplications. Hence it is possible, after a select-statement with joins to get more records that the initial table.
Define reasonable goals and test cases for the migration. Some examples:
99.9% of all contacts to be migrated correctly. With 1 M records, this means that anything lower than 1000 incorrect migrated contacts/missed, is defined as still ok.
Randomly pick 10-20 records on a base level, like 20 customer, and then compare these in the UAT/Test environment to the source system, as it is seen there. This needs to be done by the business people, so that they can have a say if the migrated data is fine.
Select some filters, like “all customers in Munich” and some other segmentations and compare source system to destination. If there are large amounts of errors, backtrack to the staging database to see where you did loose some records or created too many (not uncommon).
Complete entire transformation to destination tables in the staging db. Then you can move directly from there to dataverse. This is particularly important when moving large quantities of data when managing the data in SSIS can be problematic.
Make sure to have unique identifiers on all tables that preferably can be regenerated from the data. Store these in some “Legacy ID” field. This allows for delta-migration logic, ie. where part of the data is migrated and then the rest later. If you have some issues during one of the dataflows, and it stops on 3 230 234-th record of 6 M, you can continue from there and you don’t have to redo it all. If there is no decent way of getting a legacy id, you can generate classic row numbers by creating an identity column. This will make the migration utilize this, but only within that particular instance and load of the staging db. Hence you must be careful everytime you reload the database.
Utilize the backup-restore functionality of the dataverse environments. Do note that you can make manual backups just before you start migration. If you have a production environment, this will need to be converted to a sandbox environment before you can restore to it. Another option I got from a colleague was to use 3 different environments, with temporary names, and then just rename the final one when done.
Once you have transfered an entire table to the source system, it is typically very useful to have a mapping table, with just the table record id and the legacy id. So for instance, after migrating Contact, read all contacts from dataverse with the contactid and the legacy id. That way, when later migrating “salesorders”, which identify the customer by legacy id, it is easy to just join with this table to get the contactid.
Production environments are faster. Fastest is to ask Microsoft Support to relax throttles on all environments that are used during migration.
Use a VM that is located geographically (or really with low latency and high throughput) to where the environments are hosted. This is a very common recommendation by Kingswaysoft too.
The settings for number of threads and batch size needs to be set based on some factors, namely:
Production/Sandbox
Have throttles been relaxed
Size of payload (ie how many columns) – larger payload -> smaller batches.
Type of action – creates are faster than deletes. Updates are in-between.
I hope these tips can help you along. If you have any comments or you have other experiences in this subject, don’t hesitate to leave a comment.
During this migration and the writing of this article, I had excellent help from my highly intelligent colleague Ebba Linnea Nilsson and it is certainly true that two heads are better than one, and the end result is often a lot better than just the sum of two people. So for my final recommendation, make sure to have a good colleague with you to help you out, as you most probably will run into some issues and having someone to discuss with is really great!
As I mentioned in my previous articles, I am trying to investigate the details of how the entitlements and API Service Protections are working and are planning to be rolled out (in the case of entitlements). I had a very interesting call with some of the nice people in the product team last which shed some more light on the entitlement issue and the best practice of how they suggest the API is to be used. The suggested method is that the API request load be spread out over the different users in the instance/tenant using impersonation. I will walk through what this means and what I think about this in the article below.
First, if you have not read my previous post on entitlement, I do suggest you do this first. It describes what entitlements are compared to the API Service Protection. I still see a lot of people mixing these up and that is not strange, but they are two different aspects of this, and we need to keep track of what we are talking about.
As mentioned in that article, the point of the enacting the Entitlements, when that is coming, which still is a bit unclear, is so that the compute consumed by a small organization is proportionate compared to a large organization. So, let us go back to the actual per-user licenses and have a look at an example.
Let us say we have a 5 000 Sales Enterprise org, that means that we get:
5 000 users who each have 20 000 API request entitlements.
100 000 API Requests for non-licensed users.
Compare this to a 10 Sales Enterprise org which will have.
5 users who each have 20 000 API request entitlements.
100 000 API Requests for non-licensed users
Both these are totally independent of how many instances the first or the second org has.
The first observation is of course that the 100k API Request for non-licensed users do not scale at all with the size of the organization or the number of users. How does this then go in-line with the goal that a large org should have more compute than a small? The second observation is that 20 000 API requests, which actual also the normal UI will be using, is very large. You would have to be one busy salesperson to be able to generate 20 000 API requests manually in 24 hours, so busy I am tempted to say it is virtually impossible to break unless you have very heavy automations running under your account. This was also what the Microsoft rep I talked to mentioned, that this large number is to be used on a per user basis. Hence the natural question was, if we use impersonation in the API, will the Entitlements honor that? The answer was unequivocally: yes.
Hence, this is the clear answer on how we need to create future integrations. We need to spread the load using impersonation over many of the users in the system.
If we do this the right way, it would probably be possible for most organizations to, over time be able to build a fix for this.
However, it will not be easy as we need to have a tight control of the privileges of all the users. Let me give you an example from a customer I work with:
They are an online travel agency and have people working at the destinations with very restricted privileges. A lot of bookings (orders) are integrated from the booking systems, these should hence be spread out over many users instead of the single application user being used today. There is not natural user to direct the bookings to, as it is a B2C business, and no person at the travel agency “owns” these customers per se, so the load needs to be distributed in a more randomized fashion. So, let us say we have these users:
John Smith – System Admin (Full access)
John Doe – Power User (can create orders but not refunds)
John Surf Dude – Destination Specialist (can view but not create orders, cannot even read refunds)
When rebuilding the integration, we can use user John Smith and John Doe but not John Surf Dude and the only way of generically knowing this is checking what we want to do and comparing this to the privileges of each user to get a shortlist of users that can be used for integration.
However, we do not want to use a user that is close to 20k API requests for that day, so we might need to query the current API Request entitlement usage per user, so that we can filter the current shortlist to an even shorter list before knowing which users to use for impersonation.
Analysis
A way forward. I think this can be used, although there are some tricks to it. For my customer we might be able to cut a significant amount of API calls this way which will make a huge difference when we compared to not using this technique.
Impersonation not always viable – as in the example above, when there is not obvious owner to link to, we need to figure some other logic out of how to spread the API entitlement load. And things start to become tricky.
Impersonation is complex.
Impersonation requires more complex code. For several reasons. And the risks for errors are a lot larger as changing users is tricky. It is not strange that Microsoft still have not implemented the feature idea to be able to do user impersonation in the Power Platform UI just shows that this is less than trivial (https://powerusers.microsoft.com/t5/Power-Apps-Ideas/Allow-Admins-to-login-as-Users/idi-p/215779)
More complex dependencies on security model As mentioned above, trying to execute an action as a user that does not have the correct privileges won’t work, so we need to know that first. And setting everyone as System Administrator just will not work.
Logical user or just random users – trying to map the users to some logical connection from the other system or just randomizing the load. Logical user is probably preferrable but probably will not be a very common pattern.
Integration often system-to-system not user-to-user
Integrations are more often done on a system-to-system basis, not user-to-user basis. When looking at CRM-ERP integrations for instance, the user base of these two systems seldom overlaps except for a few users.
Takes time to refactor code to handle impersonation – There are many organizations out there with numerous complex integrations. And changing integrations on this level will require significant work to be done and the question will be if there is time to complete this work before the entitlement feature goes to GA?
Strange audit trail – if we use randomized users to update or create data in dataverse that will undoubtedly create very strange audit trails, created by and modified by fields. These are some facts that need to be taken into consideration.
Power App – per App users have very few requests – Not all licenses have 20k API requests per 24h. The Power App per App has only 1000 API Request entitlement per 24h, these can run out just by a using the system heavily. So do consider the API Entitlements when looking at the licenses.
Still not GA – Entitlements have still not gone GA. Hence the best time to let Microsoft know what you think is good or bad about this is now. But do be civil, there will be some feature like this, that will handle fairness management of compute consumption. Contact Microsoft through your local User Group, your local MVP or via the comment below or send me a message on LinkedIn and I will put you in contact with the right people. You can also submit an idea to the idea portal.
There might be a point to binding all entitlements to users, in the case that if, in the future, any overshooting would not only result in angry emails, but service degradation or shut-off for that user. Imagine having creative citizen devs creating some infinitive looping Flow or massively recursive logic unknowingly which causes a lot of requests. This approach would then just cause a block for that user, not the entire tenant. Significantly reducing the severity of the problem.
Suggestions
Personally, I think this method is just way to complex. I think just having a simple pooling on the tenant level of all the API entitlements would be fair and then deducing all usage from this. I think that Microsoft could skip the 100 000 for the non-licensed user, for simplicity. Based on the examples above, that would make:
5000 Sales Enterprise
5 000 users who each have 20 000 API request entitlements.
Total API Entitlement for the Tennant: 100 M / 24 h
5 Sales Enterprise users
5 users who each have 20 000 API request entitlements.
Total API Entitlement for the Tennant: 100 K / 24 h
And all users, and all non-licensed users use from the same pool.
As for the potential problem of creative users potentially blocking the entire tenant, I would suggest adding a “per user” API request limit, which can be changed by the admins, but by default is set at exactly the same as the entitlements. That would allow admins to reduce the limit to 10k for enterprise users, to ensure the server-to-server integrations were still enabled in a proper and entitled way.
I think this would align with Microsoft’s goals and make it easy to understand for customers and we do not have to rewrite tons of code and make strange workarounds. But maybe there is something I am missing. If so, and you see it, please leave a comment!
“Do you really need to delete records like a Ferrari?” – that question was posed to me when I, a few years ago complained about the bad performance of the Bulk Deletion functionality in Power Platform (at that time Dynamics 365 Online) to a friend at Microsoft who I will not name. And my simple answer is yes, we do need to delete records like a Ferrari, for many reasons. I will discuss why in this article and I have for that reason also created an Idea on the Power Apps Community site on this subject and I hope that you agree with me and vote for it! You will find it on the link below.
So, why is a fast bulk deletion important. I would say there are several reasons and I will walk through the ones that I have thought of, if you have any other, please drop a comment.
Keep your data in check – remove unnecessary data
GDPR and other compliancy and legal issue
Power Platform growing into Citizen developer platform
Entitlements effectivly blocks using external tools
Keeping data in check
For larger organizations, especially with many integrated modules and systems, many running Flows, workflows, Customer Voice surveys etc. the system will generate a lot of data, especially if it is a B2C scenario. A few of these have built in features that automatically remove old logs etc but most don’t and we as admins and system caretakers (isn’t it a fancy title!) need to tend to this, typically by setting up jobs that clean old data. I would recommend looking at the PPAC statistics of which tables are the largest and having a practice of doing so at regular intervals and downloading it. That way you can see the trends over time. A suggestion for an addon to the CoE Starter kit would be a trend analysis of all tables with growth numbers per week for each tables with warnings for quickly growing tables and prognosis.
However, as instances start growing over 50-100 GB in size (of structured data) it soon becomes too large to handle the data with bulk deletion. Some tables might still be managable this way, but in general the performance is has is, when I have tried to measure it (albeit a few years ago) was around 1-3 records per second. A customer I have, working with B2C for whom I wanted to remove their Voice of the Customer, which had been used a lot, had over 50 Million Survey Invites. It is not possible to remove the solution without first removing the data, and if we were to use Bulk Delete and put it on crack and it got to 10 records per second, it would still take around 2 months. I now did it with SSIS/Kingswaysoft and it took a few days. If Bulk Delete could reach around 200 records/second, it would take a little less than 3 days.
I have also noted that when trying to Bulk Delete very large datasets, Bulk Delete simply fails, as I think the FetchXML query might do a SQL Timeout or something like that. Not exactly sure what happens. As it works with Kingswaysoft I don’t know what might be the difference.
GDPR and other compliancy and legal issues
As GDPR and other similar compliancy regulations have come into play in many countries around the world, it has become ever more important to stricly follow these detailed instructions. These might be simple when you look at them on a Power Point C-level perspective but when you dig down on the detailed level, where they actually need to be implemented, things seldom are as simple as in a Power Point.
Power Platform growing into Citizen developer platform
As the Power Platform grows from being just a platform on which Dynamics 365 is delivered to being a huge platform for digitalization entire organizations with almost 100% user saturation will be coming starting to use Dataverse. The amount of data being stored in dataverse will hence grow to massive amounts and hence an effective tool to manage this data is most important. It is probably even important to such a level that Bulk Delete cannot even scratch the top of the iceberg of what we need to be able to do on a data management perspective as data will be growing and expanding in heaps and bounds and admins will not only need to manage Flows and Apps but also data in size and content.
Entitlements effectivly blocks using external tools
The soon to enacted entitlements, as mentioned in my previous post, Entitlements are not throttling | Powerplatform.se, also effectivly stop the use of external tools like SSIS/Kingswaysoft for deleting unwanted data. One of the customers I am working with generate between 10-20 M API requests PER DAY, and the bulk of these are from deletion jobs or other maintainance jobs trying to keep track of the instances. With the new entitlements charge, there is no way this can be continued, but the customer is cought between a rock and a hard place as either the data grows by leaps and bounds or the API calls becomes a huge cost and there is no easy way to handle it. What advise am I to give the customer? I would think that the most reasonable thing would be if the platform made the tools available to maintain the data to avoid the costs. If this is using bulk delete or some other more elaborate feature, that is up to the product team but I do think they should hold off on activating the entitlements until there is a good alternative for managing an instance data within the platform before this (not generating API requests).
What else is missing?
Bulk deletion is not only not being performant enough, it also lacks the effective filtering logic that is required for more complex queries. For some customers a I have had to construct rather elaborate SSIS scripts which start with a complex FetchXML and the filter the data through several Cache Transforms, for instance with GDPR consents and similar to be able to get the final list. I must admit that I havn’t tried using the new T-SQL connector for this, that it could handle the full T-SQL complexity and that it is implemented in Buld Delete or Kingswaysoft as a means to make querying more powerful.
Microsoft recently (in February) published some updates to their documentation regarding Service protection API limits or as they are sometimes referred to, throttling. Some of these, like the new recommendations on how to handle batching are rather interesting and I thought I’d give my 2 cents about this. They are also eluding a bit regarding how the network infrastructure is set up for the deployment and how to optimize when handling larger workloads using the affinity cookie setting. I did find this rather interesting too.
In short, within a 5 minute sliding window, you cannot exceed the following for one specific user.
Not more than 6 000 requests.
Not more than 1 200 (20 minutes) execution time – equal to 4 parallell processes if running at full capacity
Not more than 52 request at the same time (concurrently).
Generally, if you do not use batching, you would typically run into the first (1) or the third if running unlimited threading. If using a connection pooling with 52 connections, then you probably run into the 1:st, and if you use complex request that cause cascading behaviours or batching, then you typically run into the second (2). There are exceptions that match these. Do refer to the official docs above for details about that.
Now to the interesting part. There is a new section that is attempting to tells us how to maximize throughput. First of all, I think this is great. We really need this and we need Microsoft to tell us how to not only use the platform, but how to efficiently use their platform.
“Let the server tell you how much it can handle” This section is interesting as it recommends a rather complex approach to how to work with performance. As they further down recommend using threading, they essentially recommend building logic that dynamically increases and decreases the number of threads as the platform informs you that it has capacity. This brings me back to university math, and trying to figure out the derivative of an unknown function by sampling and finding the local max. I do however, think this is a rather tall order to recommend to your average developer. But it would be a great community tool, so feel free to build it. Consider the challenge set. Best would be if Microsoft included this in the SDK of course.
“Use multiple threads“ In this part they recommend using multiple threads. This is also my experience that this is a good idea as the processing time and latency per package causes certain delay on a per-message basis. By utilizing multi-threading with multiple connections, this overhead can be reduced. As there is a limit of 52 concurrent connections, I would recommend using a maximum of that amount of connections/threads per user.
Avoid batching Now this is really interesting. The previous recommendation was to use batching to be “nicer” to the API and get increased performance. The recommendation now is the direct opposite. This is based on the fact that the overhead in a WebAPI JSON-message is significantly smaller than that in a SOAP message and that this will reduce the difference between using batching and non-batching. They do, however, recommend using smaller batch sizes still. This is also my experience when working with Kingswaysoft. I typically (it depends on the instance and which table I am using) start with 16 threads with batches of 10 or 20. This has typically given me the best performance, with performance of +300 records/s.
There is also a comment about the fact that the using batching does not bypass the entitlement limits, ie. 20 000 API calls/24 hours for an enterprise user/100 000 API calls for all non-licensed users and so on. See more on the Entitlement limits based on which license you have here. Hence this calculation is done by after exploding the batch on the servers. This is also news to me as I previously was told that batching was exactly the way to go to limit the amount of calls.
Removing the affinity cookie – server multiplexing The details being eluded to in this section are very interesting. If I understand it correctly, the logic is as follows:
The point being that, shutting off the affinity cookie int the HttpClient will allow for more wider use of all the servers in the node (the entire setup of all the Frontends, backends, NLB etc.)
What I do wonder, is if it would be possible to store the Affinity Cookie, and hence pool it on the client side. As each time you need to hit a new front end you will loose some time while it warms up your instance, and it would hence be advantageous to be able to more tightly control this. Maybe even this could be another community tool for someone interested? I also think, I havn’t tested this, that you will get better results when working with removed affinity cookie if you do use batching, at least until all the frontends have been warmed up to your instance.
Do also note a very important sentence; “This increases throughput because limits are applied per server“. We do not know how many servers are used in a node frontend, but probably more than 10. Removing the affinity cookie could hence increase performance by at least one order of magnitute.
User multiplexing As all API limits are calculated on a per-user basis, another way to increase performance is to use what I like to call user multiplexing. This means that operations are done using several different application users at the same time. There is of course some admin work that needs to be done to set these up, and there is no OOB way of doing this but with SSIS and Kingswaysoft it is rather straight forward; just create several connections, one per user, configure them per user and then use the “Balanced Data Distributor” which can be found in the productivity pack to spread the data to different destinations that are using the different connections.
My tips My tips for getting good performance, for large scale datasets, are hence the following based on these new facts:
Continue to use batching, but don’t use huge batches. Probably around 5-20 will be ok.
Use multithreading. I typically use around 16, but that was before I knew about the removal of the affinity cookie. Hence I would recommend 16 per server. But I cannot tell you how many servers there are.
Use the remove affinity cookie setting, and if possible, figure out some way of pooling the affinity cookies instead.
Make sure your application can handle the exceptions regarding the API-limit and have some reasonable strategy for working with them. I have found that blasting the API for 5 minutes at max speed, then backing off for 5 minutes, then going full throttle again for 5 minutes, has given me better throughput overall than “being nice” and just finding the “right” speed to use to not be throttled. Not sure this strategy will work in the long run though.
Use application user multiplexing.
Suggestions to ETL vendors and others My suggestions to ETL vendors and others who build connections to Dataverse that require high performance are:
Start by visualizing the affinity cookie setting so that it is possible to set this as wanted.
Include multithreading, batching and application user multiplexing into the standard dataverse connections.
Figure out if there are an points to pooling the affinity cookies, and if so, include this into the connection.
Make the connection auto-optimize with the data it is currently sending. Ie. how many threads, size of batches, size of affinity cookie pool and number of application users to utilize.
Have different strategies for utilizing application users instead of just spreading the data evenly, it could be that one is used until it receives an exception an then it is put on hold for 5 minutes and then another is being used. Or a combination of these two if there are five application users, 3 might be used for data transfer, and two on hold in case one gets an exception and needs to be put on hold.
I hope this has given you some insights and that my 2 cents got you this far. Feel free to leave a comment if you have an questions!
Recent Comments