We experience intermittent, seemingly random brownouts of a firebase realtime database. We are beginning to shard our data into multiple databases, however, we are not sure this will solve our problem. It appears to us that firebase cannot scale to meet our needs in terms of doing frequent writes to a specific data set.
We sync data from a third-party data source in cycles (every 4-10 minutes, 1000 active jobs). Each update has the potential to change a few thousand nodes in firebase, most of which lie pretty low. However, most of the time the number of low-level nodes changed is much lower. We do differential updates on the sync'd data in order to allow very small writes to the lower-level nodes. This helps prevent our users from downloading a ton of additional data. We also batch all of our updates per cycle into only a handful of writes, between 10-20 (not sure of the performance impact of a batched write to multiple nodes vs. a write to a single node).
Here is an image of the database load graph, which includes some sharding:
Database Load
The "blue" line is our "main" database. The "orange line" is a database containing only the data that requires many writes, as described above. Currently, the main (blue) database is supporting normal operations, including reads/writes, etc.. The shard (orange) database is only handling writes. The mirror of these is pretty indicative of a "write" load issue, given that a large percentage of writes occurs in the morning.
At times, the database load reaches 100% and remains in this state for 30+ minutes.
Please let me know if I can expand on anything or explain anything in more detail. Would appreciate any suggestions on debugging strategies or explanations as to why this may be occurring.
We are actively refactoring a lot of code to mitigate this issue, however, it is not obvious what the main driver is.
Related
By using both the synchronous=OFF and journal_mode=MEMORY options, I am able to reduce the speed of updates from 15 ms to around 2 ms which is a major performance improvement. These updates happen one at a time, so many other optimizations (like using transactions about a bunch of them) are not applicable.
According to the SQLite documentation, the DB can go 'corrupt' in the worst case if there is a power outage of some type. However, is not the worst thing that can happen is for the data to be lost, or possibly part of a transaction to be lost (which I guess is a form of corruption). Is it really possible for arbitrary corruption to occur with either of these options? If so, why?
I am not using any transactions, so partially written data from transactions is not a concern, and I can handle loosing data once in a blue moon. But if 'corruption' means that all the data in the DB can be randomly changed in an unpredictable way, that would be a strong reason to not use these options.
Does any one know what the real worst-case behavior would be on iOS?
Tables are organized as B-trees with the rowid as the key.
If some writes get lost while SQLite is updating the tree structure, the entire table might become unreadable.
(The same can happen with indexes, but those could be simply dropped and recreated.)
Data is organized in pages (typically 1 KB or 4 KB). If some page update gets lost while some tree is being reorganized, all the data in these pages (i.e., some random rows from the table with nearby rowid values) might become corruped.
If SQLite needs to allocate a new page, and that page contains plausible data (e.g., deleted data from the same table), and the writing of that page gets lost, then you have incorrect data in the table, without the ability to detect it.
I have taken over a project which is a Ruby on Rails app that basically crawls the internet (in cronjobs). It crawls selected sites to build historical statistics about their subject over time.
Currently, all the raw HTML it finds is kept for future use. The HTML is saved into a Postgres database table and is currently at about 1,2 million records taking up 16 gigabytes and growing.
The application currently reads the HTML once to extract the (currently) relevant data and then flags the row as processed. However, in the future, it may be relevant to reprocess ALL files in one job to produce new information.
Since the tasks in my pipeline includes setting up backups, I don't really feel like doing daily backups of this growing amount of data, knowing that 99,9% of the backup is just static records with a practically immutable column with raw HTML.
I want this HTML out of the PostgreSQL database and into some storage. The storage must be easily manageable (inserts, reads, backups) and relatively fast. I (currently) have no requirement to search/query the content, only to resolve it by a key.
The ideal solution should allow me to insert and get data very fast, easy to backup, perhaps incrementally, perhaps support batch reads and writes for performance.
Currently, I am testing a solution where I push all these HTML files (about 15-20 KB each) to Rackspace Cloudfiles, but it takes forever and the access time is somewhat slow and relatively inconsistent (between 100 ms and several seconds). I estimate that accessing all files sequentially from rackspace could take weeks. That's not an ideal foundation for devising new statistics.
I don't believe Rackspace's CDN is the optimal solution and would like to get some ideas on how to tackle this problem in a graceful way.
I know this question has no code and is more a question about architecture and database/storage solutions, so it may be treading on the edges of SO's scope. If there's a more fitting community for the question, I'll gladly move it.
We are experiencing some serious scaling challenges for our intelligent search engine/aggregator. Our database holds around 200k objects. From profiling and newrelic it seems most of our troubles may come from the database. We are using the smallest dedicated database Heroku provide (Ronin).
We have been looking into indexing and caching. So far we managed to solve our problems by reducing database calls and caching content intelligently, but now even this seems to reach an end. We are constantly asking ourselves if our code/configuration is good enough or if we are simply not using enough "hardware".
We suspect that the database solution we buy from Heroku may be performing insufficiently. For example, just doing a simple count (no joins, no nothing) on the 200k items takes around 250ms. This seems like a long time, even though postgres is known for its bad performance on counts?
We have also started to use geolocation lookups based on latitude/longitude. Both columns are indexed floats. Doing a distance calculation involves pretty complicated math, but we are using the very well recommended geocoder gem that is suspected to run very optimized queries. Even geocoder still takes 4-10 seconds to perform a lookup on, say, 40.000 objects, returning only a limit of the first nearest 10. This again sounds like a long time, and all the experienced people we consult says that it sound very odd, again hinting at the database performance.
So basically we wonder: What can we expect from the database? Might there be a problem? And what can we expect if we decide to upgrade?
An additional question I have is: I read here that we can improve performance by loading the entire database into memory. Are we supposed to configure this ourselves and if so how?
UPDATE ON THE LAST QUESTION:
I got this from the helpful people at Heroku support:
"What this means is having enough memory (a large enough dedicated
database) to store your hot data set in memory. This isn't something
you have to do manually, Postgres is configured automatically use all
available memory on our dedicated databases.
I took a look at your database and it looks like you're currently
using about 1.25 GB of RAM, so you haven't maxed your memory usage
yet."
UPDATE ON THE NUMBERS AND FIGURES
Okay so now I've had time to look into the numbers and figures, and I'll try to answer the questions below as follows:
First of all, the db consists of around 29 tables with a lot of relations. But in reality most queries are done on a single table (some additional resources are joined in, to provide all needed information for the views).
The table has 130 columns.
Currently it holds around 200k records but only 70k are active - hence all indexes are made as partial-indexes on this "state".
All columns we search are indexed correctly and none is of text-type, and many are just booleans.
Answers to questions:
Hmm the baseline performance it's kind of hard to tell, we have sooo many different selects. The time it takes varies typically from 90ms to 250ms selecting a limit of 20 rows. We have a LOT of counts on the same table all varying from 250ms to 800ms.
Hmm well, that's hard to say cause they wont give it a shot.
We have around 8-10 users/clients running requests at the same time.
Our query load: In new relic's database reports it says this about the last 24 hours: throughput: 9.0 cpm, total time: 0.234 s, avg time: 25.9 ms
Yes we have examined the query plans of our long-running queries. The count queries are especially slow, often over 500ms for a pretty simple count on the 70k records done on indexed columns with a result around 300
I've tuned a few Rails apps hosted on Heroku, and also hosted on other platforms, and usually the problems fall into a few basic categories:
Doing too much in ruby that could be done at the db level (sorting, filtering, join data, etc)
Slow queries
Inefficient use of indexes (not enough, or too many)
Trying too hard to do it all in the db (this is not as common in rails, but does happen)
Not optimizing cacheable data
Not effectively using background processing
Right now its hard to help you because your question doesn't contain any specifics. I think you'll get a better response if you pinpoint the biggest issue you need help with and then ask.
Some info that will help us help you:
What is the average response time of your actions? (from new relic, request-log-analyzer, logs)
What is the slowest request that you want help with?
What are the queries and code in that request?
Is the site's performance different when you run it locally vs. heroku?
In the end I think you'll find that it is not an issue specific to Heroku, and if you had your app deployed on amazon, engineyard, etc you'd have the same performance. The good news is I think that your problems are common, and shouldn't be too hard to fix once you've done some benchmarking and profiling.
-John McCaffrey
We are constantly asking...
...this seems a lot...
...that is suspected...
...What can we expect...
Good news! You can put and end to seeming, suspecting wondering and expecting through the magic of measurement!!!
Seriously though, you've not mentioned any of the basic points you'd need to get a useful answer:
What's the baseline performance of the DB running a sequential scan and single-row index fetches? You say Heroku say your DB fits in RAM, so you shouldn't see disk I/O issues when you measure.
Does this performance match whatver Heroku say it should be?
How many concurrent clients?
What's your query load - what queries and how often?
Have you checked the query plans for any of your suspiciously long-running queries?
Once you've got this sort of information, maybe someone can say something useful. As it stands anything you read here is just guesswork.
First: you should check your postgres configuration. (show all from within psql or another client, or just look at postgres.conf in the data directory) The parameter with the largest impact on performance is effective_cache_size, which should be set to about (total_physical_ram - memory_in_use_by_kernel_and_all_processes). For a 4GB machine, this often is around 3GB (4-1). (this is very course tuning, but will give the best results for a first step)
Second: why do you want all the counts? Better use a typical query: just ask for what is needed, not what is available. (reason: there is no possible optimisation for a COUNT(*): eiither the whole table, or a whole index needs to be scanned)
Third: start gathering and analysing some queryplans (for typical queries that perform badly). You can get a query plan by putting EXPLAIN ANALYZE before the actual query. (another way is to increase the logging level, and obtain them from the logfile) A bad queryplan can point you at missing statistics or indexes, or even at bad data-modelling.
Newrelic monitoring can be included as an add-on for heroku (http://devcenter.heroku.com/articles/newrelic). At the very least this should give you a lot of insight into what is happening behind the scenes, and may help you pinpoint some issues.
I am having issues with a CoreData-based iOS app when it tries to build the initial DB from data sent from the server. Basically, the server sends down 1MB chunks of objects (about 3,000 per chunk), and the iOS client deserializes them and writes them into disk.
What I'm seeing is that everything is going pretty well for about the first 8 chunks (out of 44), then performance drops off dramatically and each chunk starts taking longer and longer, as in the image below. Pretty much all the time is consumed in [NSManagedObjectContext save] as you can see in the Instruments profiling data, but also it appears that the app is no longer running at 100% of CPU for some reason, like it's waiting on disk I/O or something.
A few important facts about how I'm doing this:
Each chunk is processed in its own NSManagedObjectContext with its own NSAutoreleasePool, so there is no object build-up in a non-flushed context between processing of chunks.
There is no NSUndoManager set on any of the contexts.
There is no mergeChangesFromContextDidSaveNotification: going on (i.e. the chunk contexts aren't pushing their changes into a "master" context)
I'm using a SQLite-based datastore on iOS 4.3.
The records being written do have indexes on them.
The entire sync job is processed on a single GCD background thread (i.e. dispatch_queue_create() and dispatch_async()).
I have no idea why the performance suddenly drops off like that or what can be done to address it. I have poked around and read the following, but nothing has jumped out at me yet:
http://cocoawithlove.com/2008/03/testing-core-data-with-very-big.html
Does the performance of saving a ManagedObjectContext depend on the number of contained (unchanged) objects?
Any ideas or pointers for making this app scale up to 100,000 records in the database would be much appreciated.
Edit - extra stats
This Instruments graph shows the same simulation as above (on iPad2), but includes the disk activity stats and you can see pretty plainly that all of the "not running at 100% CPU" time seems to be taken up with writing to disk.
I also ran same sync attempt running on the iOS simulator. Overall memory usage is more or less constant for each chunk except for a dictionary that contains object IDs that grows slightly over time (but these are not CoreData objects or anything that would affect saves, they are just NSNumbers). This dict is a small amount of memory compared to the total heap and so the problem is not running out of memory.
What is interesting about this test is that the CoreData Save instrument reports that the successive saves take roughly the same amount of time, which obviously conflicts with the CPU profiling information from the first set of results. It seems like CoreData thinks it is taking the same amount of time to push changes to the DB, but the DB itself (i.e. SQLite) suddenly takes a lot longer to actually stream those changes to disk.
I know this is an old issue, so this is probably no longer relevant for you, but it may be to someone else.
I've seen performance issues seeding a Core Data database over iCloud and discovered that if you have inverse relationships on the data model you can be hurt incredibly badly performance wise. The way iCloud transaction logging has been implemented, it actually seems to be an inevitable problem. Each transaction sent to iCloud (have a look at them on developer.icloud.com - they're just zipped up plists) records every relationship that is affected by a change. Unlike when you modify one end of an relationship in Core Data, and it takes care of the inverse end, the core data transaction log ends recording the changes at BOTH ends, rather than it working it out.
So if you have a 1 to many relationship, and you create another record which will end up hanging off the 'many' end - well the record at the '1' end will also be updated to reflect the fact a new additional record is now hanging off it. If you have an architecture that means you have a 'type' object that lots of 'data' objects hang off, then every time you add a new data object, the type one is going to have a transaction written for it as well - but here's the kicker, because the iCloud Core Data transactions record the ENTIRE state of edited entities, not just the changes, EVERY relationship already recorded against it is also added to the log, not just the one indicating the new subordinate record. This can quickly spiral out of control as the amount of data written grows as the number of relationships between entities grows - it ends up taking longer and longer to save batches.
I've answered a question a bit like this before here on the Apple dev forums which might be useful as I never seem to be able to describe this succinctly.
The easiest option to improve seeding performance if this scenario is what is impacting you is to switch inverse relationships off, but this isn't always an option.
More information about your implementation would help. For example, do you run this on the main thread or are you implementing background threads? However, I have seen this behavior before. When performing extensive batch operations using Core Data, it can slow down if not memory managed properly. Have you checked memory usage? Have you checked for leaks? Another thing to try is to make sure you are using NSAutoreleasePool correctly if needed. By draining the pool periodically, that may help performance.
Is it ever a good idea to let a large amount of people connect to your website while it is using sqlite?
edit: I am using it in a critical ruby on rails application that may have hundreds of concurrent users.
There are two important properties unique to SQLite that I know of that are relevant:
When doing multiple inserts, you will get better performance if you wrap them all in a single transaction. If the inserts are done individually, SQLite waits for the disk platters to rotate around completely on each insert, so that the inserted data can be read back from the disk and validated.
When writing to a SQLite file, the entire file is locked, which can cause writer starvation. This situation improved in SQLite 3.
The SQLite website says that SQLite is suitable for small to medium traffic websites, with low OLTP capability. This accounts for about 95% of all websites.