What is the average time that a rewrite API operation takes to complete rewrite for different sizes of objects?
Assume: rewrite of the same object, same source and destination but with different CMEK (of KMS).
There is no documentation regarding this. I need this information to estimate how much time a batch re-write job takes to complete when there are many objects.
I understand that latency might be a hard question to answer, but I'm only interested in whether it finishes in seconds or minutes etc(for example for an object that is couple of GBs say 3GB).
Since all of rewrite happens inside google components itself, I'm unable to figure this out. Any help or documentation related to this would be of great help.
If the source and destination buckets are in the same location and storage class, rewrite will be a fast (metadata-only) operation.
Otherwise the bytes have to be copied by the service, so time grows with object size and also will be slower if the source and destination are far apart (e.g., across an ocean). You can try doing some benchmarking experiments if you want a more precise answer than that.
Related
During my cloud dataprep adventures I have come across yet another very annoying bug.
The problem occurs when creating complex flow structures which need to be connected through reference datasets. If a certain limit is crossed in performing a number of unions or a joins with these sets, dataflow is unable to start a job.
I have had a lot of contact with support and they are working on the issue:
"Our Systems Engineer Team was able to determine the root cause resulting into the failed job. They mentioned that the job is too large. That means that the recipe (combined from all datasets) is too big, and Dataflow rejects it. Our engineering team is still investigating approaches to address this.
A workaround is to split the job into two smaller jobs. The first run the flow for the data enrichment, and then use the output as input in the other flow. While it is not ideal, this would be a working solution for the time being."
I ran into the same problem and have a fairly educated guess as to the answer. Keep in mind that DataPrep simply takes all your GUI based inputs and translates it into Apache Beam code. When you pass in a reference data set, it probably writes some AB code that turns the reference data set into a side-input (https://beam.apache.org/documentation/programming-guide/). DataFlow will perform a Parellel Do (ParDo) function where it takes each element from a PCollection, stuffs it into a worker node, and then applies the side-input data for transformation.
So I am pretty sure if the reference sets get too big (which can happen with Joins), the underlying code will take an element from dataset A, pass it to a function with side-input B...but if side-input B is very big, it won't be able to fit into the worker memory. Take a look at the Stackdriver logs for your job to investigate if this is the case. If you see 'GC (Allocation Failure)' in your logs this is a sign of not enough memory.
You can try doing this: suppose you have two CSV files to read in and process, file A is 4 GB and file B is also 4 GB. If you kick off a job to perform some type of Join, it will very quickly outgrow the worker memory and puke. If you CAN, see if you can pre-process in a way where one of the files is in the MB range and just grow the other file.
If your data structures don't lend themselves to that option, you could do what the Sys Engs suggested, split one file up into many small chunks and then feed it to the recipe iteratively against the other larger file.
Another option to test is specifying the compute type for the workers. You can iteratively grow the compute type larger and larger to see if it finally pushes through.
The other option is to code it all up yourself in Apache Beam, test locally, then port to Google Cloud DataFlow.
Hopefully these guys fix the problem soon, they don't make it easy to ask them questions, that's for sure.
I am asking a question that I assume does not have a simple black and white question but the principal of which I'm asking is clear.
Sample situation:
Lets say I have a collection of 1 million books, and I consistently want to always pull the top 100 rated.
Let's assume that I need to perform an aggregate function every time I perform this query which makes it a little expensive.
It is reasonable, that instead of running the query for every request (100-1000 a second), I would create a dedicated collection that only stores the top 100 books that gets updated every minute or so, thus instead of running a difficult query a 100 times every second, I only run it once a minute, and instead pull from a small collection of books that only holds the 100 books and that requires no query (just get everything).
That is the principal I am questioning.
Should I create a dedicated collection for EVERY query that is often
used?
Should I do it only for complicated ones?
How do I gauge which is complicated enough and which is simple enough
to leave as is?
Is there any guidelines for best practice in those types of
situations?
Is there a point where if a query runs so often and the data doesn't
change very often that I should keep the data in the server's memory
for direct access? Even if it's a lot of data? How much is too much?
Lastly,
Is there a way in MongoDB to cache results?
If so, how can I tell it to fetch the cached result, and when to regenerate the cache?
Thank you all.
Before getting to collection specifics, one does have to differentiate between "real-time data" vis-a-vis data which does not require immediate and real-time presenting of information. The rules for "real-time" systems are obviously much different.
Now to your example starting from the end. The cache of query results. The answer is not only for MongoDB. Data architects often use Redis, or memcached (or other cache systems) to hold all types of information. This though, obviously, is a function of how much memory is available to your system and the DB. You do not want to cripple the DB by giving your cache too much of available memory, and you do not want your cache to be useless by giving it too little.
In the book case, of 100 top ones, since it is certainly not a real time endeavor, it would make sense to cache the query and feed that cache out to requests. You could update the cache based upon a cron job or based upon an update flag (which you create to inform your program that the 100 have been updated) and then the system will run an $aggregate in the background.
Now to the first few points:
Should I create a dedicated collection for EVERY query that is often used?
Yes and no. It depends on the amount of data which has to be searched to $aggregate your response. And again, it also depends upon your memory limitations and btw let me add the whole server setup in terms of speed, cores and memory. MHO - cache is much better, as it avoids reading from the data all the time.
Should I do it only for complicated ones?
How do I gauge which is complicated enough and which is simple enough to leave as is?
I dont think anyone can really black and white answer to that question for your system. Is a complicated query just an $aggregate? Or is it $unwind and then a whole slew of $group etc. options following? this is really up to the dataset and how much information must actually be read and sifted and manipulated. It will effect your IO and, yes, again, the memory.
Is there a point where if a query runs so often and the data doesn't change very often that I should keep the data in the server's memory for direct access? Even if it's a lot of data? How much is too much?
See answers above this is directly connected to your other questions.
Finally:
Is there any guidelines for best practice in those types of situations?
The best you can do here is to time the procedures in your code, monitor memory usage and limits, look at the IO, study actual reads and writes on the collections.
Hope this helps.
Use a cache to store objects. For example in Redis use Redis Lists
Redis Lists are simply lists of strings, sorted by insertion order
Then set expiry to either a timeout or a specific time
Now whenever you have a miss in Redis, run the query in MongoDB and re-populate your cache. Also since cache resids in memory therefore your fetches will be extremely fast as compared to dedicated collections in MongoDB.
In addition to that, you don't have to keep have a dedicated machine, just deploy it within your application machine.
I have somewhat of a unique problem that looks similar to the problem here :
https://news.ycombinator.com/item?id=8368509
I have a high-speed traffic analysis box that is capturing at about 5 Gbps, and picking out specific packets from this to save into some format in a C++ program. Each day there will probably be 1-3 TB written to disk. Since it's network data, it's all time series down to the nanosecond level, but it would be fine to save it at second or millisecond level and have another application sort the embedded higher-resolution timestamps afterwards. My problem is deciding which format to use. My two requirements are:
Be able to write to disk at about 50 MB/s continuously with several different timestamped parameters.
Be able to export chunks of this data into MATLAB (HDF5).
Query this data once or twice a day for analytics purposes
Another nice thing that's not a hard requirement is :
There will be 4 of these boxes running independently, and it would be nice to query across all of them and combine data if possible. I should mention all 4 of these boxes are in physically different locations, so there is some overhead in sharing data.
The second one is something I cannot change because of legacy applications, but I think the first is more important. The types of queries I may want to export into matlab are something like "Pull metric X between time Y and Z", so this would eventually have to go into an HDF5 format. There is an external library called MatIO that I can use to write matlab files if needed, but it would be even better if there wasn't a translation step. I have read the entire thread mentioned above, and there are many options that appear to stand out: kdb+, Cassandra, PyTables, and OpenTSDB. All of these seem to do what I want, but I can't really figure out how easy it would be to get it into the MATLAB HDF5 format, and if any of these would make it harder than others.
If anyone has experience doing something similar, it would be a big help. Thanks!
A KDB+ tickerplant is certainly capable of capturing data at that rate, however there's lots of things you need to make sure (whatever solution you pick)
Do the machine(s) that are capturing the data have enough cores? Best to taskset a tickerplant, for example, to a core that nothing else will contend with
Similarly with disk - SSD, be sure there is no contention on the bus
Separate the workload - can write different types of data (maybe packets can be partioned by source or stream?) to different cpus/disks/tickerplant processes.
Basically there's lots of ways you can cut this. I can say though that with the appropriate hardware KDB+ could do the job. However, given you want HDF5 it's probably even better to have a simple process capturing the data and writing/converting to disk on the fly.
I have a ADO.NET/TSQL performance question. We have two options in our application:
1) One big database call with multiple result sets, then in code step through each result set and populate my objects. This results in one round trip to the database.
2) Multiple small database calls.
There is much more code reuse with Option 2 which is an advantage of that option. But I would like to get some input on what the performance cost is. Are two small round trips twice as slow as one big round trip to the database, or is it just a small, say 10% performance loss? We are using C# 3.5 and Sql Server 2008 with stored procedures and ADO.NET.
I would think it in part would depend on when you need the data. For instance if you return ten datasets in one large process, and see all ten on the screen at once, then go for it. But if you return ten datasets and the user may only click through the pages to see three of them then sending the others was a waste of server and network resources. If you return ten datasets but the user really needs to see sets seven and eight only after making changes to sets 5 and 6, then the user would see the wrong info if you returned it too soon.
If you use separate stored procs for each data set called in one master stored proc, there is no reason at all why you can't reuse the code elsewhere, so code reuse is not really an issue in my mind.
It sounds a wee bit obvious, but only send what you need in one call.
For example, we have a "getStuff" stored proc for presentation. The "updateStuff" proc calls "getStuff" proc and the client wrapper method for "updateStuff" expects type "Thing". So one round trip.
Chatty servers are one thing you prevent up front with minimal effort. Then, you can tune the DB or client code as needed... but it's hard to factor out the roundtrips later no matter how fast your code runs. In the extreme, what if your web server is in a different country to your DB server...?
Edit: it's interesting to note the SQL guys (HLGEM, astander, me) saying "one trip" and the client guys saying "multiple, code reuse"...
I am struggling with this problem myself. And I don't have an answer yet, but I do have some thoughts.
Having reviewed the answers given by others to this point, there is still a third option.
In my appllication, around ten or twelve calls are made to the server to get the data I need. Some of the datafields are varchar max and varbinary max fields (pictures, large documents, videos and sound files). All of my calls are synchronous - i.e., while the data is being requested, the user (and the client side program) has no choice but to wait. He may only want to read or view the data which only makes total sense when it is ALL there, not just partially there. The process, I believe, is slower this way and I am in the process of developing an alternative approach which is based on asynchronous calls to the server from a DLL libaray which raises events to the client to announce the progress to the client. The client is programmed to handle the DLL events and set a variable on the client side indicating chich calls have been completed. The client program can then do what it must do to prepare the data received in call #1 while the DLL is proceeding asynchronously to get the data of call #2. When the client is ready to process the data of call #2, it must check the status and wait to proceed if necessary (I am hoping this will be a short or no wait at all). In this manner, both server and client side software are getting the job done in a more efficient manner.
If you're that concerned with performance, try a test of both and see which performs better.
Personally, I prefer the second method. It makes life easier for the developers, makes code more re-usable, and modularizes things so changes down the road are easier.
I personally like option two for the reason you stated: code reuse
But consider this: for small requests the latency might be longer than what you do with the request. You have to find that right balance.
As the ADO.Net developer, your job is to make the code as correct, clear, and maintainable as possible. This means that you must separate your concerns.
It's the job of the SQL Server connection technology to make it fast.
If you implement a correct, clear, maintainable application that solves the business problems, and it turns out that the database access is the major bottleneck that prevents the system from operating within acceptable limits, then, and only then, should you start persuing ways to fix the problem. This may or may not include consolidating database queries.
Don't optimize for performance until a need arisess to do so. This means that you should analyze your anticipated use patterns and determine what the typical frequency of use for this process will be, and what user interface latency will result from the present design. If the user will receive feedback from the app is less than a few (2-3) seconds, and the application load from this process is not an inordinate load on server capacity, then don't worry about it. If otoh the user is waiting an unacceptable amount of time for a response (subjectve but definitiely measurable) or if the server is being overloaded, then it's time to begin optimization. And then, which optimization techniques will make the most sense, or be the most cost effective, depend on what your analysis of the issue tells you.
So, in the meantime, focus on maintainability. That means, in your case, code reuse
Personally I would go with 1 larger round trip.
This will definately be influenced by the exact reusability of the calling code, and how it might be refactored.
But as mentioned, this will depend on your exact situation, where maintainability vs performance could be a factor.
We used Drools as part of a solution to act as a sort of filter in a very intense processing application, maybe running up to 100 rules on 500,000 + working memory objects.
turns out that it is extremely slow.
anybody else have any experience using Drools in a batch type processing application?
Kind of depends on your rules - 500K objects is reasonable given enough memory (it has to populate a RETE network in memory, so memory usage is a multiple of 500K objects - ie space for objects + space for network structure, indexes etc) - its possible you are paging to disk which would be really slow.
Of course, if you have rules that match combinations of the same type of fact, that can cause an explosion of combinations to try, which even if you have 1 rule will be really really slow.
If you had any more information on the analysis you are doing that would probably help with possible solutions.
I've used a Drools with a stateful working memory containing over 1M facts. With some tuning of both your rules and the underlying JVM, performance can be quite good after a few minutes for initial start-up. Let me know if you want more details.
I haven't worked with the latest version of Drools (last time I used it was about a year ago), but back then our high-load benchmarks proved it to be utterly slow. A huge disappointment after having based much of our architecture on it.
At least something good I remember about drools is that their dev team was available on IRC and very helpful, you might give them a try, they're the experts after all: irc.codehaus.org #drools
I'm just learning drools myself, so maybe I'm missing something, but why is the whole batch of five hundred thousand objects added to working memory at once? The only reason I can think of is that there are rules that kick in only when two or more items in the batch are related.
If that isn't the case, then perhaps you could use a stateless session and assert one object at a time. I assume rules will run 500k times faster in that case.
Even if it is the case, do all your rules need access to all 500k objects? Could you speed things up by applying per-item rules one at a time, and then in a second phase of processing apply batch level rules using a different rulebase and working memory? This would not change the volume of data, but the RETE network would be smaller because the simple rules would have been removed.
An alternative approach would be to try and identify the related groups of objects and assert the objects in groups during the second phase, further reducing the volume of data in working memory as well as splitting up the RETE network.
Drools is not really designed to be run on a huge number of objects. It's optimized for running complex rules on a few objects.
The working memory initialization for each additional object is too slow and the caching strategies are designed to work per working memory object.
Use a stateless session and add the objects one at a time ?
I had problems with OutOfMemory errors after parsing a few thousand objects. Setting a different default optimizer solved the problem.
OptimizerFactory.setDefaultOptimizer(OptimizerFactory.SAFE_REFLECTIVE);
We were looking at drools as well, but for us the number of objects is low so this isn't an issue. I do remember reading that there are alternate versions of the same algorithm that take memory usage more into account, and are optimized for speed while still being based on the same algorithm. Not sure if any of them have made it into a real usable library though.
this optimizer can also be set by using parameter
-Dmvel2.disable.jit=true