Stream with a lot of UPDATEs and PostgreSQL - sql

I'm quite a newbie with PostgreSQL optimization and chosing whatever's appropriate job for it and whatever's not. So, I want to know whenever I'm trying to use PostgreSQL for inappropriate job, or it is suitable for it and I should set everything up properly.
Anyway, I have a need for a database with a lot of data that changes frequently.
For example, imagine an ISP, having a lot of clients, each having a session (PPP/VPN/whatever), with two self-describing frequently updated properties bytes_received and bytes_sent. There is a table with them, where each session is represented by a row with unique ID:
CREATE TABLE sessions(
id BIGSERIAL NOT NULL,
username CHARACTER VARYING(32) NOT NULL,
some_connection_data BYTEA NOT NULL,
bytes_received BIGINT NOT NULL,
bytes_sent BIGINT NOT NULL,
CONSTRAINT sessions_pkey PRIMARY KEY (id)
)
And as accounting data flows, this table receives a lot of UPDATEs like those:
-- There are *lots* of such queries!
UPDATE sessions SET bytes_received = bytes_received + 53554,
bytes_sent = bytes_sent + 30676
WHERE id = 42
When we receive a never ending stream with quite a lot (like 1-2 per second) of updates for a table with a lot (like several thousands) of sessions, probably thanks to MVCC, this makes PostgreSQL very busy. Are there any ways to speed everything up, or Postgres is just not exactly suitable for this task and I'd better consider it unsuitable for this job and put those counters to another storage like memcachedb, using Postgres only for fairly static data? But I'll miss an ability to infrequently query on this data, for example to find TOP10 downloaders, which is not really good.
Unfortunately, the amount of data cannot be lowered much. The ISP accounting example is all thought up to simplify the explanation. The real problem's with another system, which structure is somehow harder to explain.
Thanks for suggestions!

The database really isn't the best tool for collecting lots of small updates, but as I don't know your queryability and ACID requirements I can't really recommend something else. If it's an acceptable approach the application side update aggregation suggested by zzzeek can help lower the update load significantly.
There is an similar approach that can give you durability and ability to query the fresher data at some performance cost. Create a buffer table that can collect the changes to the values that need to be updated and insert the changes there. At regular intervals in a transaction rename the table to something else and create a new table in place of it. Then in a transaction aggregate all the changes, do the corresponding updates to the main table and truncate the buffer table. This way if you need a consistent and fresh snapshot of any data you can select from the main table and join in all the changes from the active and renamed buffer tables.
However if neither is acceptable you can also tune the database to deal better with heavy update loads.
To optimize the updating make sure that PostgreSQL can use heap-only tuples to store the updated versions of the rows. To do this make sure that there are no indexes on the frequently updated columns and change the fillfactor to something lower from the default 100%. You'll need to figure out a suitable fill factor on your own as it depends heavily on the details of the workload and the machine it is running on. The fillfactor needs to be low enough that allmost all of the updates fit on the same database page before autovacuum has the chance to clean up the old non-visible versions. You can tune autovacuum settings to trade off between the density of the database and vacuum overhead. Also, take into account that any long transactions, including statistical queries, will hold onto tuples that have changed after the transaction has started. See the pg_stat_user_tables view to see what to tune, especially the relationship of n_tup_hot_upd to n_tup_upd and n_live_tup to n_dead_tup.
Heavy updating will also create a heavy write ahead log (WAL) load. Tuning the WAL behavior (docs for the settings) will help lower that. In particular, a higher checkpoint_segments number and higher checkpoint_timeout can lower your IO load significantly by allowing more updates to happen in memory. See the relationship of checkpoints_timed vs. checkpoints_req in pg_stat_bgwriter to see how many checkpoints happen because either limit is reached. Raising your shared_buffers so that the working set fits in memory will also help. Check buffers_checkpoint vs. buffers_clean + buffers_backend to see how many were written to satisfy checkpoint requirements vs. just running out of memory.

You want to assemble statistical updates as they happen into an in-memory queue of some kind, or alternatively onto a message bus if you're more ambitious. A receiving process then aggregates these statistical updates on a periodic basis - which can be anywhere from every 5 seconds to every hour - depends on what you want. The counts of bytes_received and bytes_sent are then updated, with counts that may represent many individual "update" messages summed together. Additionally you should batch the update statements for multiple ids into a single transaction, ensuring that the update statements are issued in the same relative order with regards to primary key to prevent deadlocks against other transactions that might be doing the same thing.
In this way you "batch" activities into bigger chunks to control how much load is on the PG database, and also serialize many concurrent activities into a single stream (or multiple, depending on how many threads/processes are issuing updates). The tradeoff which you tune based on the "period" is, how much freshness vs. how much update load.

Related

How to optimise an SQL request containing update when?

I have an SQL request that is take more than 10 seconds to load at peak hours on my server.
UPDATE "events"
SET "metas" = 732899,
"count" = "count" + 1,
"timestamp" = 1633450429
WHERE "hash" = 'my_counter_453751'
Do you see any ways I can optimize this request?
I've tried to change my server but it doesn't change anything.
Usually, this request is almost instant, but when I have lots of users connected on my server this request takes more than 10 seconds and I don't understand why.
Any advice would help!
UPDATE "events" SET ... WHERE "hash" = 'my_counter_453751'
For this to work efficiently, you need an index on hash (if at all possible, it would be somewhat better to retrieve an integer-type primary index).
You also need as few other indexes as possible, because an index on metas, for example, too would be updated if you changed that column, and that takes time.
The fact that under load the update time goes up in that way, though, is suspicious. One of the few things that could explain this is disk thrashing. With many users, the events table gets evicted from memory and requires reloading. With too many indexes, and even more if there is no index on hash at all, this can be calamitous.
So I would also review the memory assigned to the RDBMS server, and the size of the events table. You could trim it down, or benefit from sharding. Or you could split a historical table into several tables, one per month - a poor man's sharding. This complicates retrieving information from that table, but accelerates writes.
I have also reaped huge benefits from dirty tricks such as writing these updates to a different, smaller table (or even to a temporary file). So, none of my hundred simultaneous clients has ever to do anything to the huge table; they all access the tiny one. Then a periodic cron job would run every minute or so, and perform all the updates together, loading the huge table in memory only once - and not locking or interfering in any way with the other threads, then remove the processed records from the tiny table. Maybe some such strategy can be doable for you too.
For a historic log table, you could write to a daily log and then pour the daily log into the large table late every night.

Is there any perfomance issues when inserting into a large SQL Server table which is being queried?

I use SQL Server. I got a large table - millions of rows. And I iterate through them (SELECT .. WHERE ..). This is a long operation (and I assume can't be shorter).
So what am I asking is if there will be any problems to insert data into that table in the progress of selecting? If yes, what should I do to reduce that? Same questing for update command (with indexed parameters of course).
Yes, you will have performance, and more specifically, locking and blocking issues. If your SELECT statements are using indexes, which they should be, these indexes will be locked every time that you INSERT data into the table. Since the table is relatively large, the lock will probably be long enough to block your SELECT statements, and deadlocks are likely as well.
This might be a scenario where you need to re-evaluate your table structure, and possibly even consider denormalizing to avoid this.
You might also consider Enabling Row Versioning-Based Isolation Levels, assuming that you can throughly test the rest of your system to understand the impact.
The answer is yes, absolutely. A simple solution (if it's an acceptable trade off within your application) is to specify the NOLOCK locking hint. IE:
select * from table with NOLOCK
The tradeoff is that you won't get a consistent read, but in many cases this isn't problem.
It's generally not a good idea to have long running queries on a database with frequent updates. This decrease performance significantly because of locking.
It might be a good idea to look into data warehouses and see if that is something that you could use. That would enable you to have the transactions on a separate database and the bulk load from it in to another database that would have your warehouse.
This would greatly improve performance for both inserts and queries. The trans-actional database could have no indexes, and the warehouse could have all the indexes you want.
You could also put the warehouse in a column store database. That would give you the best query time with the minimal effort because there isn't any need to create indexes in a column store, all you would have to do is to design the schema properly. The drawback with column stores is how ever that inserts, updates and deletes are very slow compared to relational databases. But bulk loading from the transactional database should do the trick. If you require the data to be very up to date, you could bulk load every few minutes. If you just need data from the previous day you could bulk load into the warehouse each night.
The possibilities are endless. If you want to look into column store warehouses you could try MonetDB. Its an open source column store so you could try it out and see if that's anything that suits you.
Do not assume execution time can't be shorter. If you query a date range, an index on date is a must!
Solve your problem indexing on date field:
-- please use correct names for your_table and date_field --
CREATE INDEX index_name ON your_table date_field
Warehousing, as per #Gisli, is a good option: build a copy of the data elsewhere, and run your long-running queries there, freeing up the "main" database for OLTP processing.
If this is not an option, you can mess around with snapshot isolation (something I know about, but have never worked with personally). Esssentially, this will take a "snapshot" of the database at the point in time you start the query, and will execute the query as if no subsequent changes were made to the database, even if changes are made to the database while the query is running. More importantly, any such changes are "real" and permanent. Think of it like a short-term branching of your database.
The duration of the branch (snapshot) is where I get weak. I believe you can have the snapshot last for the duration of the query, which means you'd (possibly) never be able to get the same results for a given query twice (if the data changes while you are running it); or you can create a "saved" snapshot that can be re-used over and over until you get around to deleting it. Be wary with this, you don't want your system to get cluttered up with old forgotten branches of past data!
There is no PROBLEM. SQL Serve is built to deal with this kind of situations, you just need to set the correct isolation level on the transactions.
There are several possible scenarios, for example, if you don't mind reading the data that is being inserted, set your isolation level to read uncommited on your read transaction. If you are inserting values in a range and reading values on another range, you can use SERIALIZABLE.
Take a look at the possible isolation levels:
http://msdn.microsoft.com/en-us/library/ms173763.aspx

Running Updates on a large, heavily used table

I have a large table (~170 million rows, 2 nvarchar and 7 int columns) in SQL Server 2005 that is constantly being inserted into. Everything works ok with it from a performance perspective, but every once in a while I have to update a set of rows in the table which causes problems. It works fine if I update a small set of data, but if I have to update a set of 40,000 records or so it takes around 3 minutes and blocks on the table which causes problems since the inserts start failing.
If I just run a select to get back the data that needs to be updated I get back the 40k records in about 2 seconds. It's just the updates that take forever. This is reflected in the execution plan for the update where the clustered index update takes up 90% of the cost and the index seek and top operator to get the rows take up 10% of the cost. The column I'm updating is not part of any index key, so it's not like it reorganizing anything.
Does anyone have any ideas on how this could be sped up? My thought now is to write a service that will just see when these updates have to happen, pull back the records that have to be updated, and then loop through and update them one by one. This will satisfy my business needs but it's another module to maintain and I would love if I could fix this from just a DBA side of things.
Thanks for any thoughts!
Actually it might reorganise pages if you update the nvarchar columns.
Depending on what the update does to these columns they might cause the record to grow bigger than the space reserved for it before the update.
(See explanation now nvarchar is stored at http://www.databasejournal.com/features/mssql/physical-database-design-consideration.html.)
So say a record has a string of 20 characters saved in the nvarchar - this takes 20*2+2(2 for the pointer) bytes in space. This is written at the initial insert into your table (based on the index structure). SQL Server will only use as much space as your nvarchar really takes.
Now comes the update and inserts a string of 40 characters. And oops, the space for the record within your leaf structure of your index is suddenly too small. So off goes the record to a different physical place with a pointer in the old place pointing to the actual place of the updated record.
This then causes your index to go stale and because the whole physical structure requires changing you see a lot of index work going on behind the scenes. Very likely causing an exclusive table lock escalation.
Not sure how best to deal with this. Personally if possible I take an exclusive table lock, drop the index, do the updates, reindex. Because your updates sometimes cause the index to go stale this might be the fastest option. However this requires a maintenance window.
You should batch up your update into several updates (say 10000 at a time, TEST!) rather than one large one of 40k rows.
This way you will avoid a table lock, SQL Server will only take out 5000 locks (page or row) before esclating to a table lock and even this is not very predictable (memory pressure etc). Smaller updates made in this fasion will at least avoid concurrency issues you are experiencing.
You can batch the updates using a service or firehose cursor.
Read this for more info:
http://msdn.microsoft.com/en-us/library/ms184286.aspx
Hope this helps
Robert
The mos brute-force (and simplest) way is to have a basic service, as you mentioned. That has the advantage of being able to scale with the load on the server and/or the data load.
For example, if you have a set of updates that must happen ASAP, then you could turn up the batch size. Conversely, for less important updates, you could have the update "server" slow down if each update is taking "too long" to relieve some of the pressure on the DB.
This sort of "heartbeat" process is rather common in systems and can be very powerful in the right situations.
Its wired that your analyzer is saying it take time to update the clustered Index . Did the size of the data change when you update ? Seems like the varchar is driving the data to be re-organized which might need updates to index pointers(As KMB as already pointed out) . In that case you might want to increase the % free sizes on the data and the index pages so that the data and the index pages can grow without relinking/reallocation . Since update is an IO intensive operation ( unlike read , which can be buffered ) the performance also depends on several factors
1) Are your tables partitioned by data 2) Does the entire table lies in the same SAN disk ( Or is the SAN striped well ?) 3) How verbose is the transaction logging . Can the buffer size of the transaction loggin increased to support larger writes to the log to suport massive inserts ?
Its also important which API/Language are you using? e.g JDBC support a batch update feature which makes the updates a little bit efficient if you are doing multiple updates .

Do relational databases provide a feasible backend for a process historian?

In the process industry, lots of data is read, often at a high frequency, from several different data sources, such as NIR instruments as well as common instruments for pH, temperature, and pressure measurements. This data is often stored in a process historian, usually for a long time.
Due to this, process historians have different requirements than relational databases. Most queries to a process historian require either time stamps or time ranges to operate on, as well as a set of variables of interest.
Frequent and many INSERT, many SELECT, few or no UPDATE, almost no DELETE.
Q1. Is relational databases a good backend for a process historian?
A very naive implementation of a process historian in SQL could be something like this.
+------------------------------------------------+
| Variable |
+------------------------------------------------+
| Id : integer primary key |
| Name : nvarchar(32) |
+------------------------------------------------+
+------------------------------------------------+
| Data |
+------------------------------------------------+
| Id : integer primary key |
| Time : datetime |
| VariableId : integer foreign key (Variable.Id) |
| Value : float |
+------------------------------------------------+
This structure is very simple, but probably slow for normal process historian operations, as it lacks "sufficient" indexes.
But for example if the Variable table would consist of 1.000 rows (rather optimistic number), and data for all these 1.000 variables would be sampled once per minute (also an optimistic number) then the Data table would grow with 1.440.000 rows per day. Lets continue the example, estimate that each row would take about 16 bytes, which gives roughly 23 megabytes per day, not counting additional space for indexes and other overhead.
23 megabytes as such perhaps isn't that much but keep in mind that numbers of variables and samples in the example were optimistic and that the system will need to be operational 24/7/365.
Of course, archiving and compression comes to mind.
Q2. Is there a better way to accomplish this? Perhaps using some other table structure?
I work with a SQL Server 2008 database that has similar characteristics; heavy on insertion and selection, light on update/delete. About 100,000 "nodes" all sampling at least once per hour. And there's a twist; all of the incoming data for each "node" needs to be correlated against the history and used for validation, forecasting, etc. Oh, there's another twist; the data needs to be represented in 4 different ways, so there are essentially 4 different copies of this data, none of which can be derived from any of the other data with reasonable accuracy and within reasonable time. 23 megabytes would be a cakewalk; we're talking hundreds-of-gigabytes to terabytes here.
You'll learn a lot about scale in the process, about what techniques work and what don't, but modern SQL databases are definitely up to the task. This system that I just described? It's running on a 5-year-old IBM xSeries with 2 GB of RAM and a RAID 5 array, and it performs admirably, nobody has to wait more than a few seconds for even the most complex queries.
You'll need to optimize, of course. You'll need to denormalize frequently, and maintain pre-computed aggregates (or a data warehouse) if that's part of your reporting requirement. You might need to think outside the box a little: for example, we use a number of custom CLR types for raw data storage and CLR aggregates/functions for some of the more unusual transactional reports. SQL Server and other DB engines might not offer everything you need up-front, but you can work around their limitations.
You'll also want to cache - heavily. Maintain hourly, daily, weekly summaries. Invest in a front-end server with plenty of memory and cache as many reports as you can. This is in addition to whatever data warehousing solution you come up with if applicable.
One of the things you'll probably want to get rid of is that "Id" key in your hypothetical Data table. My guess is that Data is a leaf table - it usually is in these scenarios - and this makes it one of the few situations where I'd recommend a natural key over a surrogate. The same variable probably can't generate duplicate rows for the same timestamp, so all you really need is the variable and timestamp as your primary key. As the table gets larger and larger, having a separate index on variable and timestamp (which of course needs to be covering) is going to waste enormous amounts of space - 20, 50, 100 GB, easily. And of course every INSERT now needs to update two or more indexes.
I really believe that an RDBMS (or SQL database, if you prefer) is as capable for this task as any other if you exercise sufficient care and planning in your design. If you just start slinging tables together without any regard for performance or scale, then of course you will get into trouble later, and when the database is several hundred GB it will be difficult to dig yourself out of that hole.
But is it feasible? Absolutely. Monitor the performance constantly and over time you will learn what optimizations you need to make.
It sounds like you're talking about telemetry data (time stamps, data points).
We don't use SQL databases for this (although we do use SQL databases to organize it); instead, we use binary streaming files to capture the actual data. There are a number of binary file formats that are suitable for this, including HDF5 and CDF. The file format we use here is a proprietary compressible format. But then, we deal with hundreds of megabytes of telemetry data in one go.
You might find this article interesting (links directly to Microsoft Word document):
http://www.microsoft.com/caseStudies/ServeFileResource.aspx?4000003362
It is a case study from the McClaren group, describing how SQL Server 2008 is used to capture and process telemetry data from formula one race cars. Note that they don't actually store the telemetry data in the database; instead, it is stored in the file system, and the FILESTREAM capability of SQL Server 2008 is used to access it.
I believe you're headed in the right path. We have a similar situation were we work. Data comes from various transport / automation systems across various technologies such as manufacturing, auto, etc. Mainly we deal with the big 3: Ford, Chrysler, GM. But we've had a lot of data coming in from customers like CAT.
We ended up extracting data into a database and as long as you properly index your table, keep updates to a minimum and schedule maintenance (rebuild indexes, purge old data, update statistics) then I see no reason for this to be a bad solution; in fact I think it is a good solution.
Certainly a relational database is suitable for mining the data after the fact.
Various nuclear and particle physics experiments I have been involved with have explored several points from not using a RDBMS at all though storing just the run summaries or the run summaries and the slowly varying environmental conditions in the DB all the way to cramming every bit collected into the DB (though it was staged to disk first).
When and where the data rate allows more and more groups are moving towards putting as much data as possible into the database.
IBM Informix Dynamic Server (IDS) has a TimeSeries DataBlade and RealTime Loader which might provide relevant functionality.
Your naïve schema records each reading 100% independently, which makes it hard to correlate across readings- both for the same variable at different times and for different variables at (approximately) the same time. That may be necessary, but it makes life harder when dealing with subsequent processing. How much of an issue that is depends on how often you will need to run correlations across all 1000 variables (or even a significant percentage of the 1000 variables, where significant might be as small as 1% and would almost certainly start by 10%).
I would look to combine key variables into groups that can be recorded jointly. For example, if you have a monitor unit that records temperature, pressure and acidity (pH) at one location, and there are perhaps a hundred of these monitors in the plant that is being monitored, I would expect to group the three readings plus the location ID (or monitor ID) and time into a single row:
CREATE TABLE MonitorReading
(
MonitorID INTEGER NOT NULL REFERENCES MonitorUnit,
Time DATETIME NOT NULL,
PhReading FLOAT NOT NULL,
Pressure FLOAT NOT NULL,
Temperature FLOAT NOT NULL,
PRIMARY KEY (MonitorID, Time)
);
This saves having to do self-joins to see what the three readings were at a particular location at a particular time, and uses about 20 bytes instead of 3 * 16 = 48 bytes per row. If you are adamant that you need a unique ID integer for the record, that increases to 24 or 28 bytes (depending on whether you use a 4-byte or 8-byte integer for the ID column).
Yes, a DBMS is appropriate for this, although not the fastest option. You will need to invest in a reasonable system to handle the load though. I will address the rest of my answer to this problem.
It depends on how beefy a system you're willing to throw at the problem. There are two main limiters for how fast you can insert data into a DB: bulk I/O speed and seek time. A well-designed relational DB will perform at least 2 seeks per insertion: one to begin the transaction (in case the transaction can not be completed), and one when the transaction is committed. Add to this additional storage to seek to your index entries and update them.
If your data are large, then the limiting factor will be how fast you can write data. For a hard drive, this will be about 60-120 MB/s. For a solid state disk, you can expect upwards of 200 MB/s. You will (of course) want extra disks for a RAID array. The pertinent figure is storage bandwidth AKA sequential I/O speed.
If writing a lot of small transactions, the limitation will be how fast your disk can seek to a spot and write a small piece of data, measured in IO per second (IOPS). We can estimate that it will take 4-8 seeks per transaction (a reasonable case with transactions enabled and an index or two, plus some integrity checks). For a hard drive, the seek time will be several milliseconds, depending on disk RPM. This will limit you to several hundred writes per second. For a solid state disk, the seek time is under 1 ms, so you can write several THOUSAND transactions per second.
When updating indices, you will need to do about O(log n) small seeks to find where to update, so the DB will slow down as the record counts grow. Remember that a DB may not write in the most efficient format possible, so data size may be bigger than you expect.
So, in general, YES, you can do this with a DBMS, although you will want to invest in good storage to ensure it can keep up with your insertion rate. If you wish to cut on cost, you may want to roll data over a specific age (say 1 year) into a secondary, compressed archive format.
EDIT:
A DBMS is probably the easiest system to work with for storing recent data, but you should strongly consider the HDF5/CDF format someone else suggested for storing older, archived data. It is an flexible and widely supported format, provides compression, and provides for compression and VERY efficient storage of large time series and multi-dimensional arrays. I believe it also provides for some methods of indexing in the data. You should be able to write a little code to fetch from these archive files if data is too old to be in the DB.
There is probably a data structure that would be more optimal for your given case than a relational database.
Having said that, there are many reasons to go with a relational DB including robust code support, backup & replication technology and a large community of experts.
Your use case is similar to high-volume financial applications and telco applications. Both are frequently inserting data and frequently doing queries that are both time-based and include other select factors.
I worked on a mid-sized billing project that handled cable bills for millions of subscribers. That meant an average of around 5 rows per subscriber times a few million subscribers per month in the financial transaction table alone. That was easily handled by a mid-size Oracle server using (now) 4 year old hardware and software. Large billing platforms can have 10x that many records per unit time.
Properly architected and with the right hardware, this case can be handled well by modern relational DB's.
Years ago, a customer of ours tried to load an RDBMS with real-time data collected from monitoring plant machinery. It didn't work in a simplistic way.
Is relational databases a good backend for a process historian?
Yes, but. It needs to store summary data, not details.
You'll need a front-end based in-memory and on flat files. Periodic summaries and digests can be loaded into an RDBMS for further analysis.
You'll want to look at Data Warehousing techniques for this. Most of what you want to do is to split your data into two essential parts ---
Facts. The data that has units. Actual measurements.
Dimensions. The various attributes of the facts -- date, location, device, etc.
This leads you to a more sophisticated data model.
Fact: Key, Measure 1, Measure 2, ..., Measure n, Date, Geography, Device, Product Line, Customer, etc.
Dimension 1 (Date/Time): Year, Quarter, Month, Week, Day, Hour
Dimension 2 (Geography): location hierarchy of some kind
Dimension 3 (Device): attributes of the device
Dimension *n*: attributes of each dimension of the fact
You may want to look at KDB. It is specificaly optimized for this kind of usage: many inserts, few or no updates or deletes.
It isn't as easy to use as traditional RDBMS though.
The other aspect to consider is what kind of selects you're doing. Relational/SQL databases are great for doing complex joins dependent on multiple indexes, etc. They really can't be beaten for that. But if you're not doing that kind of thing, they're probably not such a great match.
If all you're doing is storing per-time records, I'd be tempted to roll your own file format ... even just output the stuff as CSV (groans from the audience, I know, but it's hard to beat for wide acceptance)
It really depends on your indexing/lookup requirements, and your willingness to write tools to do it.
You may want to take a look at a Stream Data Manager System (SDMS).
While not addressing all your needs (long-time persistence), sliding windows over time and rows and frequently changing data are their points of strength.
Some useful links:
Stanford Stream Data Manager
Stream Mill
Material about Continuous Queries
AFAIK major database makers all should have some kind of prototype version of an SDMS in the works, so I think it's a paradigm worth checking out.
I know you're asking about relational database systems, but those are unicorns. SQL DBMSs are probably a bad match for your needs because no current SQL system (I know of) provides reasonable facilities to deal with temporal data. depending on your needs you might or might not have another option in specialized tools and formats, see e. g. rrdtool.

SQL Server Efficiently dropping a group of rows with millions and millions of rows

I recently asked this question:
MS SQL share identity seed amongst tables
(Many people wondered why)
I have the following layout of a table:
Table: Stars
starId bigint
categoryId bigint
starname varchar(200)
But my problem is that I have millions and millions of rows. So when I want to delete stars from the table Stars it is too intense on SQL Server.
I cannot use built in partitioning for 2005+ because I do not have an enterprise license.
When I do delete though, I always delete a whole category Id at a time.
I thought of doing a design like this:
Table: Star_1
starId bigint
CategoryId bigint constaint rock=1
starname varchar(200)
Table: Star_2
starId bigint
CategoryId bigint constaint rock=2
starname varchar(200)
In this way I can delete a whole category and hence millions of rows in O(1) by doing a simple drop table.
My question is, is it a problem to have hundreds of thousands of tables in your SQL Server? The drop in O(1) is extremely desirable to me. Maybe there's a completely different solution I'm not thinking of?
Edit:
Is a star ever modified once it is inserted? No.
Do you ever have to query across star categories? I never have to query across star categories.
If you are looking for data on a particular star, would you know which table to query? Yes
When entering data, how will the application decide which table to put the data into? The insertion of star data is done all at once at the start when the categoryId is created.
How many categories will there be? You can assume there will be infinite star categories. Let's say up to 100 star categories per day and up to 30 star categories not needed per day.
Truly do you need to delete the whole category or only the star that the data changed for? Yes the whole star category.
Have you tried deleting in batches? Yes we do that today, but it is not good enough.
od enough.
Another technique is mark the record for deletion? There is no need to mark a star as deleted because we know the whole star category is eligible to be deleted.
What proportion of them never get used? Typically we keep each star category data for a couple weeks but sometimes need to keep more.
When you decide one is useful is that good for ever or might it still need to be deleted later?
Not forever, but until a manual request to delete the category is issued.
If so what % of the time does that happen? Not that often.
What kind of disc arrangement are you using? Single filegroup storage and no partitioning currently.
Can you use sql enterprise ? No. There are many people that run this software and they only have sql standard. It is outside of their budget to get ms sql enterprise.
My question is, is it a problem to have hundreds of thousands of tables in your SQL Server?
Yes. It is a huge problem to have this many tables in your SQL Server. Every object has to be tracked by SQL Server as metadata, and once you include indexes, referential constraints, primary keys, defaults, and so on, then you are talking about millions of database objects.
While SQL Server may theoretically be able to handle 232 objects, rest assured that it will start buckling under the load much sooner than that.
And if the database doesn't collapse, your developers and IT staff almost certainly will. I get nervous when I see more than a thousand tables or so; show me a database with hundreds of thousands and I will run away screaming.
Creating hundreds of thousands of tables as a poor-man's partitioning strategy will eliminate your ability to do any of the following:
Write efficient queries (how do you SELECT multiple categories?)
Maintain unique identities (as you've already discovered)
Maintain referential integrity (unless you like managing 300,000 foreign keys)
Perform ranged updates
Write clean application code
Maintain any sort of history
Enforce proper security (it seems evident that users would have to be able to initiate these create/drops - very dangerous)
Cache properly - 100,000 tables means 100,000 different execution plans all competing for the same memory, which you likely don't have enough of;
Hire a DBA (because rest assured, they will quit as soon as they see your database).
On the other hand, it's not a problem at all to have hundreds of thousands of rows, or even millions of rows, in a single table - that's the way SQL Server and other SQL RDBMSes were designed to be used and they are very well-optimized for this case.
The drop in O(1) is extremely desirable to me. Maybe there's a completely different solution I'm not thinking of?
The typical solution to performance problems in databases is, in order of preference:
Run a profiler to determine what the slowest parts of the query are;
Improve the query, if possible (i.e. by eliminating non-sargable predicates);
Normalize or add indexes to eliminate those bottlenecks;
Denormalize when necessary (not generally applicable to deletes);
If cascade constraints or triggers are involved, disable those for the duration of the transaction and blow out the cascades manually.
But the reality here is that you don't need a "solution."
"Millions and millions of rows" is not a lot in a SQL Server database. It is very quick to delete a few thousand rows from a table of millions by simply indexing on the column you wish to delete from - in this case CategoryID. SQL Server can do this without breaking a sweat.
In fact, deletions normally have an O(M log N) complexity (N = number of rows, M = number of rows to delete). In order to achieve an O(1) deletion time, you'd be sacrificing almost every benefit that SQL Server provides in the first place.
O(M log N) may not be as fast as O(1), but the kind of slowdowns you're talking about (several minutes to delete) must have a secondary cause. The numbers do not add up, and to demonstrate this, I've gone ahead and produced a benchmark:
Table Schema:
CREATE TABLE Stars
(
StarID int NOT NULL IDENTITY(1, 1)
CONSTRAINT PK_Stars PRIMARY KEY CLUSTERED,
CategoryID smallint NOT NULL,
StarName varchar(200)
)
CREATE INDEX IX_Stars_Category
ON Stars (CategoryID)
Note that this schema is not even really optimized for DELETE operations, it's a fairly run-of-the-mill table schema you might see in SQL server. If this table has no relationships, then we don't need the surrogate key or clustered index (or we could put the clustered index on the category). I'll come back to that later.
Sample Data:
This will populate the table with 10 million rows, using 500 categories (i.e. a cardinality of 1:20,000 per category). You can tweak the parameters to change the amount of data and/or cardinality.
SET NOCOUNT ON
DECLARE
#BatchSize int,
#BatchNum int,
#BatchCount int,
#StatusMsg nvarchar(100)
SET #BatchSize = 1000
SET #BatchCount = 10000
SET #BatchNum = 1
WHILE (#BatchNum <= #BatchCount)
BEGIN
SET #StatusMsg =
N'Inserting rows - batch #' + CAST(#BatchNum AS nvarchar(5))
RAISERROR(#StatusMsg, 0, 1) WITH NOWAIT
INSERT Stars2 (CategoryID, StarName)
SELECT
v.number % 500,
CAST(RAND() * v.number AS varchar(200))
FROM master.dbo.spt_values v
WHERE v.type = 'P'
AND v.number >= 1
AND v.number <= #BatchSize
SET #BatchNum = #BatchNum + 1
END
Profile Script
The simplest of them all...
DELETE FROM Stars
WHERE CategoryID = 50
Results:
This was tested on an 5-year old workstation machine running, IIRC, a 32-bit dual-core AMD Athlon and a cheap 7200 RPM SATA drive.
I ran the test 10 times using different CategoryIDs. The slowest time (cold cache) was about 5 seconds. The fastest time was 1 second.
Perhaps not as fast as simply dropping the table, but nowhere near the multi-minute deletion times you mentioned. And remember, this isn't even on a decent machine!
But we can do better...
Everything about your question implies that this data isn't related. If you don't have relations, you don't need the surrogate key, and can get rid of one of the indexes, moving the clustered index to the CategoryID column.
Now, as a rule, clustered indexes on non-unique/non-sequential columns are not a good practice. But we're just benchmarking here, so we'll do it anyway:
CREATE TABLE Stars
(
CategoryID smallint NOT NULL,
StarName varchar(200)
)
CREATE CLUSTERED INDEX IX_Stars_Category
ON Stars (CategoryID)
Run the same test data generator on this (incurring a mind-boggling number of page splits) and the same deletion took an average of just 62 milliseconds, and 190 from a cold cache (outlier). And for reference, if the index is made nonclustered (no clustered index at all) then the delete time only goes up to an average of 606 ms.
Conclusion:
If you're seeing delete times of several minutes - or even several seconds then something is very, very wrong.
Possible factors are:
Statistics aren't up to date (shouldn't be an issue here, but if it is, just run sp_updatestats);
Lack of indexing (although, curiously, removing the IX_Stars_Category index in the first example actually leads to a faster overall delete, because the clustered index scan is faster than the nonclustered index delete);
Improperly-chosen data types. If you only have millions of rows, as opposed to billions, then you do not need a bigint on the StarID. You definitely don't need it on the CategoryID - if you have fewer than 32,768 categories then you can even do with a smallint. Every byte of unnecessary data in each row adds an I/O cost.
Lock contention. Maybe the problem isn't actually delete speed at all; maybe some other script or process is holding locks on Star rows and the DELETE just sits around waiting for them to let go.
Extremely poor hardware. I was able to run this without any problems on a pretty lousy machine, but if you're running this database on a '90s-era Presario or some similar machine that's preposterously unsuitable for hosting an instance of SQL Server, and it's heavily-loaded, then you're obviously going to run into problems.
Very expensive foreign keys, triggers, constraints, or other database objects which you haven't included in your example, which might be adding a high cost. Your execution plan should clearly show this (in the optimized example above, it's just a single Clustered Index Delete).
I honestly cannot think of any other possibilities. Deletes in SQL Server just aren't that slow.
If you're able to run these benchmarks and see roughly the same performance I saw (or better), then it means the problem is with your database design and optimization strategy, not with SQL Server or the asymptotic complexity of deletions. I would suggest, as a starting point, to read a little about optimization:
SQL Server Optimization Tips (Database Journal)
SQL Server Optimization (MSDN)
Improving SQL Server Performance (MSDN)
SQL Server Query Processing Team Blog
SQL Server Performance (particularly their tips on indexes)
If this still doesn't help you, then I can offer the following additional suggestions:
Upgrade to SQL Server 2008, which gives you a myriad of compression options that can vastly improve I/O performance;
Consider pre-compressing the per-category Star data into a compact serialized list (using the BinaryWriter class in .NET), and store it in a varbinary column. This way you can have one row per category. This violates 1NF rules, but since you don't seem to be doing anything with individual Star data from within the database anyway anyway, I doubt you'd be losing much.
Consider using a non-relational database or storage format, such as db4o or Cassandra. Instead of implementing a known database anti-pattern (the infamous "data dump"), use a tool that is actually designed for that kind of storage and access pattern.
Must you delete them? Often it is better to just set an IsDeleted bit column to 1, and then do the actual deletion asynchronously during off hours.
Edit:
This is a shot in the dark, but adding a clustered index on CategoryId may speed up deletes. It may also impact other queries adversely. Is this something you can test?
This was the old technique in SQL 2000 , partitioned views and remains a valid option for SQL 2005. The problem does come in from having large quantity of tables and the maintenance overheads associated with them.
As you say, partitioning is an enterprise feature, but is designed for this large scale data removal / rolling window effect.
One other option would be running batched deletes to avoid creating 1 very large transaction, creating hundreds of far smaller transactions, to avoid lock escalations and keep each transaction small.
Having separate tables is partitioning - you are just managing it manually and do not get any management assistance or unified access (without a view or partitioned view).
Is the cost of Enterprise Edition more expensive than the cost of separately building and maintaining a partitioning scheme?
Alternatives to the long-running delete also include populating a replacement table with identical schema and simply excluding the rows to be deleted and then swapping the table out with sp_rename.
I'm not understanding why whole categories of stars are being deleted on a regular basis? Presumably you are having new categories created all the time, which means your number of categories must be huge and partitioning on (manually or not) that would be very intensive.
Maybe on the Stars table set the PK to non-clustered and add a clustered index on categoryid.
Other than that, is the server setup well done regarding best practices for performance? That is using separate physical disks for data and logs, not using RAID5, etc.
When you say deleting millions of rows is "too intense for SQL server", what do you mean? Do you mean that the log file grows too much during the delete?
All you should have to do is execute the delete in batches of a fixed size:
DECLARE #i INT
SET #i = 1
WHILE #i > 0
BEGIN
DELETE TOP 10000 FROM dbo.SuperBigTable
WHERE CategoryID = 743
SELECT #i = ##ROWCOUNT
END
If your database is in full recovery mode, you will have to run frequent transaction log backups during this process so that it can reuse the space in the log. If the database is in simple mode, you shouldn't have to do anything.
My only other recommendation is to make sure that you have an appropriate index in CategoryId. I might even recommend that this be the clustered index.
If you want to optimize on a category delete clustered composite index with category at the first place might do more good than damage.
Also you could describe the relationships on the table.
It sounds like the transaction log is struggling with the size of the delete. The transaction log grows in units, and this takes time whilst it allocates more disk space.
It is not possible to delete rows from a table without enlisting a transaction, although it is possible to truncate a table using the TRUNCATE command. However this will remove all rows in the table without condition.
I can offer the following suggestions:
Switch to a non-transactional database or possibly flat files. It doesn't sound like you need atomicity of a transactional database.
Attempt the following. After every x deletes (depending on size) issue the following statement
BACKUP LOG WITH TRUNCATE_ONLY;
This simply truncates the transaction log, the space remains for the log to refill. However Im not sure howmuch time this will add to the operation.
What do you do with the star data? If you only look at data for one category at any given time this might work, but it is hard to maintain. Every time you have a new category, you will have to build a new table. If you want to query across categories, it becomes more complex and possibly more expensive in terms of time. If you do this and do want to query across categories a view is probably best (but do not pile views on top of views). If you are looking for data on a particular star, would you know which table to query? If not then how are you going to determine which table or are you goign to query them all? When entering data, how will the application decide which table to put the data into? How many categories will there be? And incidentally relating to each having a separate id, use the bigint identities and combine the identity with the category type for your unique identifier.
Truly do you need to delete the whole category or only the star that the data changed for?
And do you need to delete at all, maybe you only need to update information.
Have you tried deleting in batches (1000 records or so at a time in a loop). This is often much faster than deleting a million records in one delete statement. It often keeps the table from getting locked during the delete as well.
Another technique is mark the record for deletion. Then you can run a batch process when usage is low to delete those records and your queries can run on a view that excludes the records marked for deletion.
Given your answers, I think your proposal may be reasonable.
I know this is a bit of a tangent, but is SQL Server (or any relational database) really a good tool for this job? What relation database features are you actually using?
If you are dropping whole categories at a time, you can't have much referential integrity depending on it. The data is read only, so you don't need ACID for data updates.
Sounds to me like you are using basic SELECT query features?
Just taking your idea of many tables - how can you realise that...
What about using dynamic queries.
create the table of categories that have identity category_id column.
create the trigger on insert for this tale - in it create table for stars with the name dynamically made from category_id.
create the trigger on delete - in it drop the corresponding stars table also with the help of dynamically created sql.
to select stars of concrete category you can use function that returns table. It will take category_id as a parameter and return result also through dynamic query.
to insert stars of new category you firstly insert new row in categories table and then insert stars to appropriate table.
Another direction in which I would make some researches is using xml typed column for storing stars data. The main idea here is if you need to operate stars only by categories than why not to store all stars of concrete category in one cell of the table in xml format. Unfortunately I absolutely cannot imaging what will be the performance of such decision.
Both this variants are just like ideas in brainstorm.
As Cade pointed out, adding a table for each category is manually partitioning the data, without the benefits of the unified access.
There will never be any deletions for millions of rows that happen as fast as dropping a table, without the use of partitions.
Therefore, it seems like using a separate table for each category may be a valid solution. However, since you've stated that some of these categories are kept, and some are deleted, here is a solution:
Create a new stars table for each new
category.
Wait for the time period to expire where you decide whether the stars for the category are kept or not.
Roll the records into the main stars table if you plan on keeping them.
Drop the table.
This way, you will have a finite number of tables, depending on the rate you add categories and the time period where you decide if you want them or not.
Ultimately, for the categories that you keep, you're doubling the work, but the extra work is distributed over time. Inserts to the end of the clustered index may be experienced less by the users than deletes from the middle. However, for those categories that you're not keeping, you're saving tons of time.
Even if you're not technically saving work, perception is often the bigger issue.
I didn't get an answer to my comment on the original post, so I am going under some assumptions...
Here's my idea: use multiple databases, one for each category.
You can use the managed ESE database that ships with every version of Windows, for free.
Use the PersistentDictionary object, and keep track of the starid, starname pairs that way. If you need to delete a category, just delete the PersistentDictionary object for that category.
PersistentDictionary<int, string> starsForCategory = new PersistentDictionary<int, string>("Category1");
This will create a database called "Category1", on which you can use standard .NET dictionary methods (add, exists, foreach, etc).