I have a table with more than a millon rows. This table is used to index tiff images. Each image has fields like date, number, etc. I have users that index these images in batches of 500. I need to know if it is better to first insert 500 rows and then perform 500 updates or, when the user finishes indexing, to do the 500 inserts with all the data. A very important thing is that if I do the 500 inserts at first, this time is free for me because I can do it the night before.
So the question is: is it better to do inserts or inserts and updates, and why? I have defined a id value for each image, and I also have other indices on the fields.
Updates in Sql server result in ghosted rows - i.e. Sql crosses one row out and puts a new one in. The crossed out row is deleted later.
Both inserts and updates can cause page-splits in this way, they both effectively 'add' data, it's just that updates flag the old stuff out first.
On top of this updates need to look up the row first, which for lots of data can take longer than the update.
Inserts will just about always be quicker, especially if they are either in order or if the underlying table doesn't have a clustered index.
When inserting larger amounts of data into a table look at the current indexes - they can take a while to change and build. Adding values in the middle of an index is always slower.
You can think of it like appending to an address book: Mr Z can just be added to the last page, while you'll have to find space in the middle for Mr M.
Doing the inserts first and then the updates does seem to be a better idea for several reasons. You will be inserting at a time of low transaction volume. Since inserts have more data, this is a better time to do it.
Since you are using an id value (which is presumably indexed) for updates, the overhead of updates will be very low. You would also have less data during your updates.
You could also turn off transactions at the batch (500 inserts/updates) level and use it for each individual record, thus reducing some overhead.
Finally, test this out to see the actual performance on your server before making a final decision.
This isn't a cut and dry question. Krishna's and Galegian's points are spot on.
For updates, the impact will be lessened if the updates are affecting fixed-length fields. If updating varchar or blob fields, you may add a cost of page splits during update when the new value surpasses the length of the old value.
I think inserts will run faster. They do not require a lookup (when you do an update you are basically doing the equivalent of a select with the where clause). And also, an insert won't lock the rows the way an update will, so it won't interfere with any selects that are happening against the table at the same time.
The execution plan for each query will tell you which one should be more expensive. The real limiting factor will be the writes to disk, so you may need to run some tests while running perfmon to see which query causes more writes and causes the disk queue to get the longest (longer is bad).
I'm not a database guy, but I imagine doing the inserts in one shot would be faster because the updates require a lookup whereas the inserts do not.
Related
We are trying to implement table partitioning for a Data Warehouse Fact table which contains approximately 400M rows. Our ETL takes data from source system 50 days backwards (new rows, modified rows, based on source system timestamp) from the previous load. So in every ETL cycle there are new rows coming in, and also old rows which are updating the corresponding rows in the Fact table. The idea is to insert new rows into the Fact table and update modified rows.
The partition column would be date (int, YYYYMMDD) and we are considering to partition by month.
As far as I'm concerned, table partitioning would ease our inserts via fast partition switch operations. We could split the most recent partition to create a new free partition, load new rows into a staging table (using date constraint, e.g for the most recent month) and then use partition switch operation to "move" new rows into the partitioned Fact table. But how can we handle the modified rows which should update the corresponding rows in the Fact table? Those rows can contain data from the previous month(s). Does partition switch help here? Usually INSERT and UPDATE rows are determined by an ETL tool (e.g. SSIS in our case) or by MERGE statement. How partitioning works in these kind of situations?
I'd take another look at the design and try to figure out if there's a way around the updates. Here are a few implications of updating the fact table:
Performance: Updates are fully logged transactions. Big fact tables also have lots of data to read and write.
Cubes: Updating the fact table requires reprocessing the affected partitions. As your fact table continues to grow, the cube processing time will continue to as well.
Budget: Fast storage is expensive. Updating big fact tables will require lots of fast reads and writes.
Purist theory: You should not change the fact table unless the initial value was an error (ie the user entered $15,000 instead of $1,500). Any non-error scenario will be changing the originally recorded transaction.
What is changing? Are the changing pieces really attributes of a dimension? If so, can they be moved to a dimension and have changes handled with a Slowly Changing Dimension type task?
Another possibility, can this be accomplished via offsetting transactions? Example:
The initial InvoiceAmount was $10.00. Accounting later added $1.25 for tax then billed the customer for $11.25. Rather than updating the value to $11.25, insert a record for $1.25. The sum amount for the invoice will still be $11.25 and you can do a minimally logged insert rather than a fully logged update to accomplish.
Not only is updating the fact table a bad idea in theory, it gets very expensive and non-scalable as the fact table grows. You'll be reading and writing more data, requiring more IOPS from the storage subsytem. When you get ready to do analytics, cube processing will then throw in more problems.
You'll also have to constantly justify to management why you need so many IOPS for the data warehouse. Is there business value/justification in needing all of those IOPS for your constant changing "fact" table?
If you can't find a way around updates on the fact table, at least establish a cut-off point where the data is determined read-only. Otherwise, you'll never be able to scale.
Switching does not help here.
Maybe you can execute updates concurrently using multiple threads on distinct ranges of rows. That might speed it up. Be careful not to trigger lock escalation so you get good concurrency.
Also make sure that you update the rows mostly in ascending sort order of the clustered index. This helps with disk IO (this technique might not work well with multi-threading).
There are as many reasons to update a fact record as there are non-identifying attributes in the fact. Unless you plan on a "first delete" then "insert", you simply cannot avoid updates. You cannot simply say "record the metric deltas as new facts".
Short version of the question:
If you have a table with a large number of small rows and you want to retrieve a single record from this table via an index probably consisting of two columns is this likely to be something that wil be low cost and fast or high cost and slow
Longer version of question and background:
I am a consultant working with a software development company and I have an argument with them about the performance implications of a piece of functionality that I want to add to the application they are building (and I am designing).
At the moment, we write out a log record every time somebody retrieves a client record. I want to put the name and time of the last person prevously to access that record onto the client page each time that record is retrieved.
They are saying that the performance implications of this will be high but based on my reasonable but not expert knowledge of how B trees work, this doesn't seem right even if the table is very large.
If you create an index on the GUID of the client record and the date/time of access (descending), then you ought to be able to retrieve the required record via an index scan which would just need to find the first entry for that GUID and then stop? And that with a b-tree index, most of the index would be cached so the number of physical disc accesses needed would be very small and the query time therefore significantly less than 1s.
Or have I got this completely wrong
You will have problems with GUID index fragmentation but because your rows do not increase in size (as you said in the comments) you will not have page-splitting problems. The random insert issue is fixable by doing reorganizing and rebuilding.
Besides that, there is nothing wrong with your approach. If the table is larger than RAM you will likely have a single disk IO per access (the intermediate index levels will be cached). If your data fits in RAM you will pay about 0.2 to 0.5ms per query. If your data is on a magnetic disk a seek will likely require 8-12ms. On an SSD you are back to 0.2ms to 0.5ms (maybe 0.05ms more).
Why don't you just create some test data (by selecting a cross product from sys.object of 1M rows) and measure it. It takes little time and you will find out for sure.
should be low cost and fast since the columns are indexed and that would be O(n) I think
You say last person to access? You mean that for every read you will have a write?
And that write is going to change an indexed date time column?
Then I would be worried too.
Writing on each record read will cause you lots of extra disk writes. This will block reads and it might be bad to your caching too. You also need to update your index a lot, and since you change the indexed data your index will be very fragmented.
It depends.
A single retrieval will be low cost and fast
on a decent indexed table
running on decent hardware
over a decent network
On the other hand, it takes time nonetheless.
If we are talking about one retrieval per hour, don't sweat over it. If we are talking about thousands of retrievals per second (as opposed to currently none) it will start to add up to the point it would be noticable.
Some questions you need to adress
Is my hardware up to spec
Does adding two fields result in a page split (unlikely)
How many extra pages need to be read for your regular result sets
How many retrievals/sec will be made
How many inserts/sec (triggering an index update) will be made
After you've adressed these questions, you should be able to make the determination yourself. As far as my gut feelings go, I would be surprised you would notice the performance difference.
I have noticed an interesting performance change that happens around 1,5 million entered values. Can someone give me a good explanation why this is happening?
Table is very simple. It is consisted of (bigint, bigint, bigint, bool, varbinary(max))
I have a pk clusered index on first three bigints. I insert only boolean "true" as data varbinary(max).
From that point on, performance seems pretty constant.
Legend: Y (Time in ms) | X (Inserts 10K)
I am also curios about constant relatively small (sometimes very large) spikes I have on the graph.
Actual Execution Plan from before spikes.
Legend:
Table I am inserting into: TSMDataTable
1. BigInt DataNodeID - fk
2. BigInt TS - main timestapm
3. BigInt CTS - modification timestamp
4. Bit: ICT - keeps record of last inserted value (increases read performance)
5. Data: Data
Bool value Current time stampl keeps
Enviorment
It is local.
It is not sharing any resources.
It is fixed size database (enough so it does not expand).
(Computer, 4 core, 8GB, 7200rps, Win 7).
(Sql Server 2008 R2 DC, Processor Affinity (core 1,2), 3GB, )
Have you checked the execution plan once the time goes up? The plan may change depending on statistics. Since your data grow fast, stats will change and that may trigger a different execution plan.
Nested loops are good for small amounts of data, but as you can see, the time grows with volume. The SQL query optimizer then probably switches to a hash or merge plan which is consistent for large volumes of data.
To confirm this theory quickly, try to disable statistics auto update and run your test again. You should not see the "bump" then.
EDIT: Since Falcon confirmed that performance changed due to statistics we can work out the next steps.
I guess you do a one by one insert, correct? In that case (if you cannot insert bulk) you'll be much better off inserting into a heap work table, then in regular intervals, move the rows in bulk into the target table. This is because for each inserted row, SQL has to check for key duplicates, foreign keys and other checks and sort and split pages all the time. If you can afford postponing these checks for a little later, you'll get a superb insert performance I think.
I used this method for metrics logging. Logging would go into a plain heap table with no indexes, no foreign keys, no checks. Every ten minutes, I create a new table of this kind, then with two "sp_rename"s within a transaction (swift swap) I make the full table available for processing and the new table takes the logging. Then you have the comfort of doing all the checking, sorting, splitting only once, in bulk.
Apart from this, I'm not sure how to improve your situation. You certainly need to update statistics regularly as that is a key to a good performance in general.
Might try using a single column identity clustered key and an additional unique index on those three columns, but I'm doubtful it would help much.
Might try padding the indexes - if your inserted data are not sequential. This would eliminate excessive page splitting and shuffling and fragmentation. You'll need to maintain the padding regularly which may require an off-time.
Might try to give it a HW upgrade. You'll need to figure out which component is the bottleneck. It may be the CPU or the disk - my favourite in this case. Memory not likely imho if you have one by one inserts. It should be easy then, if it's not the CPU (the line hanging on top of the graph) then it's most likely your IO holding you back. Try some better controller, better cached and faster disk...
I am developing a Job application which executes multiple parallel jobs. Every job will pull data from third party source and process. Minimum records are 100,000. So i am creating new table for each job (like Job123. 123 is jobId) and processing it. When job starts it will clear old records and get new records and process. Now the problem is I have 1000 jobs and the DB has 1000 tables. The DB size is drastically increased due to lots of tables.
My question is whether it is ok to create new table for each job. or have only one table called Job and have column jobId, then enter data and process it. Only problem is every job will have 100,000+ records. If we have only one table, whether DB performance will be affected?
Please let me know which approach is better.
Don't create all those tables! Even though it might work, there's a huge performance hit.
Having a big table is fine, that's what databases are for. But...I suspect that you don't need 100 million persistent records, do you? It looks like you only process one Job at a time, but it's unclear.
Edit
The database will grow to the largest size needed, but the space from deleted records is reused. If you add 100k records and delete them, over and over, the database won't keep growing. But even after the delete it will take up as much space as 100k records.
I recommend a single large table for all jobs. There should be one table for each kind of thing, not one table for each thing.
If you make the Job ID the first field in the clustered index, SQL Server will use a b-tree index to determine the physical order of data in the table. In principle, the data will automatically be physically grouped by Job ID due to the physical sort order. This may not stay strictly true forever due to fragmentation, but that would affect a multiple table design as well.
The performance impact of making the Job ID the first key field of a large table should be negligible for single-job operations as opposed to having a separate table for each job.
Also, a single large table will generally be more space efficient than multiple tables for the same amount of total data. This will improve performance by reducing pressure on the cache.
I have a sproc that puts 750K records into a temp table through a query as one of its first actions. If I create indexes on the temp table before filling it, the item takes about twice as long to run compared to when I index after filling the table. (The index is an integer in a single column, the table being indexed is just two columns each a single integer.)
This seems a little off to me, but then I don't have the firmest understanding of what goes on under the hood. Does anyone have an answer for this?
If you create a clustered index, it affects the way the data is physically ordered on the disk. It's better to add the index after the fact and let the database engine reorder the rows when it knows how the data is distributed.
For example, let's say you needed to build a brick wall with numbered bricks so that those with the highest number are at the bottom of the wall. It would be a difficult task if you were just handed the bricks in random order, one at a time - you wouldn't know which bricks were going to turn out to be the highest numbered, and you'd have to tear the wall down and rebuild it over and over. It would be a lot easier to handle that task if you had all the bricks lined up in front of you, and could organize your work.
That's how it is for the database engine - if you let it know about the whole job, it can be much more efficient than if you just feed it a row at a time.
It's because the database server has to do calculations each and every time you insert a new row. Basically, you end up reindexing the table each time. It doesn't seem like a very expensive operation, and it's not, but when you do that many of them together, you start to see the impact. That's why you usually want to index after you've populated your rows, since it will just be a one-time cost.
Think of it this way.
Given
unorderedList = {5, 1,3}
orderedList = {1,3,5}
add 2 to both lists.
unorderedList = {5, 1,3,2}
orderedList = {1,2,3,5}
What list do you think is easier to add to?
Btw ordering your input before load will give you a boost.
You should NEVER EVER create an index on an empty table if you are going to massively load it right afterwards.
Indexes have to be maintained as the data on the table changes, so imagine as if for every insert on the table the index was being recalculated (which is an expensive operation).
Load the table first and create the index after finishing with the load.
That's were the performance difference is going.
After performing large data manipulation operations, you frequently have to update the underlying indexes. You can do that by using the UPDATE STATISTICS [table] statement.
The other option is to drop and recreate the index which, if you are doing large data insertions, will likely perform the inserts much faster. You can even incorporate that into your stored procedure.
this is because if the data you insert is not in the order of the index, SQL will have to split pages to make room for additional rows to keep them together logically
This due to the fact that when SQL Server indexes table with data it is able to produce exact statistics of values in indexed column. At some moments SQL Server will recalculate statistics, but when you perform massive inserts the distribution of values may change after the statistics was calculated last time.
The fact that statistics is out of date can be discovered on Query Analyzer. When you see that on a certain table scan number of rows expected differs to much from actual numbers of rows processed.
You should use UPDATE STATISTICS to recalculate distribution of values after you insert all the data. After that no performance difference should be observed.
If you have an index on a table, as you add data to the table SQL Server will have to re-order the table to make room in the appropriate place for the new records. If you're adding a lot of data, it will have to reorder it over and over again. By creating an index only after the data is loaded, the re-order only needs to happen once.
Of course, if you are importing the records in index order it shouldn't matter so much.
In addition to the index overhead, running each query as a transaction is a bad idea for the same reason. If you run chunks of inserts (say 100) within 1 explicit transaction, you should also see a performance increase.