|
Sander Rossel wrote: Bit of mansplaining there 
Please accept my apologies.
I know that you are a developer of some years experience, but you pressed my "teaching" button, and I tend to over-explain at times. 
Freedom is the freedom to say that two plus two make four. If that is granted, all else follows.
-- 6079 Smith W.
|
|
|
|
|
Daniel Pfeffer wrote: and I tend to over-explain at times
Better than the opposite, which I tend to do at times.
|
|
|
|
|
A few other ideas:
* Check for index fragmentation. If it's high that will hurt performance. In SSMS, right click on Indexes and select either Rebuild or Reindex to see what the values are. Then OK to do that action.
* Even if you need 20 columns to make a unique index, then do it as a composite key. That will still perform better than you manually doing the check. Manual checks may also have race conditions between the check and the insert.
* When creating indexes, don't forget about included columns. These are columns that are not a part of the index, but are retrieved with it. It allows for your index to be small and fast, but you get the data you need faster.
* Use the Execution Plan in SSMS to see where your bottlenecks are on the database side. Sometimes it will also offer index suggestions.
Enjoy!
Bond
Keep all things as simple as possible, but no simpler. -said someone, somewhere
|
|
|
|
|
|
That's impressive 
|
|
|
|
|
Never tried it, as we don't use SQL Server anymore, but it looks very promising 
|
|
|
|
|
I'd like to know more technical details. Assuming 2500 lines, 40 values per line, these values are text? numeric? Just how large is the overall file? What's the connection speed between the web site and the azure based system? Does your azure system have sufficient resources?
Since you have a lot of error handling per your own admission  I'd think you could add some logging into the mix to see where you are spending your time.
But this: "Last week, it took 17 tries to import one file" is a smoking gun. Solve the timeout issue, and I'd give it 50/50 your performance issues go away. You're not moving that much data. One other suggestion, double the resources (temporarily) for the azure system to make sure it's not under-resourced.
Off the cuff thoughts.
Charlie Gilley
“They who can give up essential liberty to obtain a little temporary safety deserve neither liberty nor safety.” BF, 1759
Has never been more appropriate.
|
|
|
|
|
charlieg wrote: these values are text? numeric?
Text, numbers, decimals, dates.
charlieg wrote: Just how large is the overall file?
I have an initial file of 248 KB.
charlieg wrote: What's the connection speed between the web site and the azure based system?
They both run in Azure in the same region, so I suspect the connection is fast.
charlieg wrote: Does your azure system have sufficient resources?
Yeah, we don't have the fastest database (50 DTUs), but it's plenty sufficient for everything else.
charlieg wrote: see where you are spending your time.
Inserting 2500 lines.
await context.BulkInsertAsync(lines); is the exact line
Although I suspect there may be some other long running queries in production.
Possibly getting all the lines for a specific date.
I'm looking at my new code now (and added an index) and the inserting takes the longest by far (a rewrite of this functionality was necessary for other reasons too, I just gave it more priority because of this).
charlieg wrote: Solve the timeout issue, and I'd give it 50/50 your performance issues go away.
The timeout issue is the performance issue 
|
|
|
|
|
I suspect your problem is that Entity Framework tries to be "intelligent".
My first guess is that turning off AutoDetectChangesEnabled would solve your problem.
But seriously, I would skip the whole BulkInsert thingy and go directly to SqlBulkCopy instead.
|
|
|
|
|
Jörgen Andersson wrote: I suspect your problem is that Entity Framework tries to be "intelligent".
Yeah, using "vanilla" EF takes minutes to insert 2500 rows, so that's not an option.
I'm using the EFCore.BulkExtensions library for this one.
Jörgen Andersson wrote: and go directly to SqlBulkCopy instead
Wouldn't I have to insert before I can copy? 
|
|
|
|
|
No, you just need to read the CSV-file to an IEnumerable<t> of sorts and connect it to an EntityDataReader that you use as an input to SqlBulkCopy.
EntityDatareader is a part of System.Data.EntityClient.
Or you can use a CSV-Reader[^] that you connect directly to SqlBulkCopy.
|
|
|
|
|
The BulkInsert is using SqlBulkCopy internally.
Using SqlBulkCopy directly is about equally fast.
|
|
|
|
|
Sander Rossel wrote: Yeah, we don't have the fastest database (50 DTUs), but it's plenty sufficient for everything else.
I ran some testing of our application with Azure SQL some time ago and found it to be very poor for an OLAP style IO bound workload, and probably checking a big index for uniqueness is similar. 50 DTU's on Basic or Standard tier is no more than 200 IOPS - DTU-based purchasing model - Azure SQL Database | Microsoft Learn[^]
For us it was better to use a SQL server on an Azure virtual machine, then we could do optimisations like striping the database across multiple smaller (unmanaged) hdd's using Windows Storage Spaces. Nowadays there are probably better managed disk options with SSD's etc to get you a decent IO performance level. Configure storage for SQL Server VMs - SQL Server on Azure VMs | Microsoft Learn[^]
|
|
|
|
|
50 DTU's on Standard tier is more than enough for everything else we have.
It's not an OLAP application, it's more administrative.
It has some master data, you can import and edit orders (it's this importing that's causing me headaches) and some one-off reports and other functionality.
The import lines (about 2500 a file) are grouped, some prices are calculated and those prices are added per group in code, which is super fast.
Other than that, the lines are printed on a report.
All in all it's a pretty small application that ties the production and the financial software together.
The database currently stores 3.38 GB of data of which this import table is about 2/3rds (and that's all we do with those lines) 
|
|
|
|
|
Inspired by this question, I went to take a more detailed look at where the time to load a 1.2 million line, 0.5 gigabyte, CSV file goes in a particular application that I'm working on. That process takes about 5 seconds too, by some coincidence.
There are three steps: reading the file (all at once), converting the encoding, and parsing the CSV into objects.
- reading the file (File.ReadAllBytes) takes about 150 milliseconds. (so we're reading the file at 3.3GB/s, which is reasonable, near the maximum sequential read speed of the SSD)
- converting the encoding (aka Encoding.GetString) takes about 900 milliseconds. (at 5GHz, this comes down to 9 cycles per byte, should it take 9 cycles per byte to do encoding conversion? sounds slow to me)
- parsing / creating objects takes the rest of the time. (this code is crap and takes at least 10 times as long as it should)
There's no database involved. Loading 2k rows in the same time basically comes down to saying that using a database makes loading the data over 500 times as slow - compared to code that is slow already. Is that reasonable? Databases are infamous for being pure molasses, but that sounds extreme.
On the other hand, it's only 5 seconds, not a big deal.
|
|
|
|
|
harold aptroot wrote: load a 1.2 million line, 0.5 gigabyte, CSV file
harold aptroot wrote: Databases are infamous for being pure molasses, but that sounds extreme.
Well, at 2500 lines and 40 values per line, it comes down to storing 100,000 values in the correct place.
Storing the CSV directly takes a few 100 milliseconds.
The database actions are the only slow lines of code here.
Reading the CSV and putting it to objects is milliseconds (and that's using a general CSV parser that uses reflection).
harold aptroot wrote: using a database makes loading the data over 500 times as slow
I guess that sounds about right.
Depending on how fast your DB server is, of course
|
|
|
|
|
What David said. I have similar requirements for updating a huge table (or set of tables) with a relatively small CSV dataset. I put the CSV into a separate table, do the data validation (like duplicate checks) with SQL and indexed columns on the criteria for duplicates, and perform the resulting updates/inserts in C# and email the customer with the change log. Very very fast.
|
|
|
|
|
How do you handle validation towards your users?
A user imports a file and wants to know why it failed (e.g. some value is invalid or the file was already (partly) imported).
|
|
|
|
|
Well, since these updaters run as an automated process, the customer gets an email.
Internally, all errors, warnings, and performed actions are logged, so if it were a manually initiated process, we could provide a full report to the user on the website as well.
|
|
|
|
|
Your answer is right here on CP: A Fast CSV Reader[^]
That article is not written by me, but I've used this component since sometime way back in 2008 or so.
It has been updated along the way.
That is a fantastic and amazingly fast (see perf stats in the article).
It's super easy to use and will solve your problems fast. I remember getting it to work in about 15 minutes. Import the component into your project & the API calls are intuitive.
Seriously try it and I'm sure it'll make you happy. 
Let me know what you think.
Here's a snippet of the performance from the article:
From article: To give more down-to-earth numbers, with a 45 MB CSV file containing 145 fields and 50,000 records, the reader was processing about 30 MB/sec. So all in all, it took 1.5 seconds! The machine specs were P4 3.0 GHz, 1024 MB.
|
|
|
|
|
I'm pretty sure it's a great tool, but it's not the CSV I'm having trouble with.
I get all my values in objects in mere milliseconds with my own generic CSV reader.
Besides, this project uses .NET Framework 2.0, which is like 15 versions, two complete overhauls and twenty years behind my version of .NET 
My problem is getting all those values in a database in an acceptable time frame (and 30+ seconds is not acceptable  ).
|
|
|
|
|
Have a session when they are doing the process and see why the time out occurs. What changed ? Was it working ok before? Why is it taking 17 tries now? maybe u can do a sql monitor and see the iops/sessions etc....
Caveat Emptor.
"Progress doesn't come from early risers – progress is made by lazy men looking for easier ways to do things." Lazarus Long
|
|
|
|
|
This is a nightmare scenario: Quote: Of course it wasn't always like that, but the import table in Azure SQL has over 2 million records now and I need to validate for double records,
I've done this before and it's slow no matter what I do and our db is very optimized. What I wound up doing was pulling and storing in an array, all of the unique fields only. It's one pull (huge, but with only one or two values per record, doable). Then in the double-record validation, I just check each new record against that one array, in advance of any db insert/update work, sorting the new records into two arrays, one for insert, one for update. That way the inserts are quicker. If you are trashing the duplicates, you can skip the update part.
|
|
|
|
|
So, you've isolated it to the DB load. In ghost jobs of long past, I know inserting records into tables with indexes could bite you, since the db has to do it's thing. The application was a "real-time" database that gated prepaid phone calls. Although we didn't do batch operations, we might have anywhere from 512-4096 call handler threads all hitting the database relatively at the same time (don't want to get pedantic) . However, most of these transactions were either reads or updates. Creation occurred only for new customers.
So, lots of index use, but no new index entries.
Please add a reminder to your task list to update the thread with what the actual problem is and how you solved it
Oh, one other suggestion - it sounds like to me you could collect these files and set up an offline test system to play - remove the web complexity. I know you've done this a long time, so you probably already are doing this.
Charlie Gilley
“They who can give up essential liberty to obtain a little temporary safety deserve neither liberty nor safety.” BF, 1759
Has never been more appropriate.
|
|
|
|
|
Interesting. We have issues with CSV upload but it is in the upload failing itself, which is being used due to the lack of an API from an outside vendor. It's AWS to local. That's a topic in itself with a long, dumb winding road with a large vendor. Haven't had issues with Azure connects, though.
My first question is process: It sounds like you have control over both ends of this? If so, why a CSV rather than a direct upsert? Five seconds feels like an eternity (it does really feel that way today!) but CSVs to DB with validation takes a while. Our environment is different, using Postgres on a nix with a triggered background cron and our line count is only in the hundreds per upload, so numbers don't compare. What's an import "table" and why does it exist, or is that just wording for a real destination table?
This caught my eye: Quote: Not so much an issue in my development environment, but apparently a bottleneck in production.
You mean the time just isn't an issue or does it run faster? Not sure why 5 seconds would create a "bottleneck" in production with five uploads a day? The timeout is generated on your end, right? Bump it up or calc it on estimated process time. If it ultimately works, let it run, no? Yeah, it's annoying looking at a spinner for a bit, but even if it's 30 seconds five times a day, no one is that busy.
Sorry, haven't worked with SQL Server in years but we do find the more you can put into a longer SQL query, the better the performance vs outside processing. Again, with SQL Server caveat above, I'd also index the heck out of it with 2 mil records.
|
|
|
|
|
General 
News 
Suggestion 
Question 
Bug 
Answer 
Joke 
Praise 
Rant 
Admin 
Submit an article or tip
