I'm really confused about how or what AWS services to use for my case.
I have a web application which stores user interaction events. Currently these events are stored on a RDS table. Each event contains about 6 fields like timestamp, event type, userID, pageID, etc etc. Currently I have millions of event records on each account schema. When I try to generate reports out of this raw data - the reports are extremely slow since I do complex aggregation queries over long time period. a report of a time period of 30 days might take 4 minutes to generate on RDS.
Is there any way to make these reports running MUCH faster? I was thinking about storing the events on DynamoDB, but I cannot run such complex queries on the data, and to do any attribute based sorting.
Is there a good service combination to achieve this? Maybe using RedShift, EMP, Kinesis?
I think Redshift is your solution.
I'm working with a dataset that generates about 2.000.000 new rows each day and I made really complex operations on it. You could take advance of Redshift sort keys, and order your data by date.
Also if you do complex aggregate functions I really recommend to denormalize all the information and insert it in only one table with all the data. Redshift uses a very efficient, and automatic, column compression you won't have problems with the size of the dataset.
My usual solution to problems like this is to have a set of routines that rollup and store the aggregated results, to various levels in additional RDS tables. This transactional information you are storing isn't likely to change once logged, so, for example, if you find yourself running daily/weekly/monthly rollups of various slices of data, run the query and store those results, not necessarily at the final level that you will need, but at a level that significantly reduces the # of rows that goes into those eventual rollups. For example, have a daily table that summarizes eventtype, userid and pageId one row per day, instead of one row per event (or one row per hour instead of day) - you'll need to figure out the most logical rollups to make, but you get the idea - the goal is to pre-summarize at the levels that will reduce the amount of raw data, but still gives you plenty of flexibility to serve your reports.
You can always go back to the granular/transactional data as long as you keep it around, but there is not much to be gained by constantly calculating the same results every time you want to use the data.
Related
I have been working on AWS Redshift and kind of curious about which of the data loading (full reload) method is more performant.
Approach 1 (Using Truncate):
Truncate the existing table
Load the data using Insert Into Select statement
Approach 2 (Using Drop and Create):
Drop the existing table
Load the data using Create Table As Select statement
We have been using both in our ETL, but I am interested in understanding what's happening behind the scene on AWS side.
In my opinion - Drop and Create Table As statement should be more performant as it reduces the overhead of scanning/handling associated data blocks for table needed in Insert Into statement.
Moreover, truncate in AWS Redshift does not reseed identity columns - Redshift Truncate table and reset Identity?
Please share your thoughts.
Redshift operates on 1MB blocks as the base unit of storage and coherency. When changes are made to a table it is these blocks that are "published" for all to see when the changes are committed. A table is just a list (data structure) of block ids that compose it and since there can be many versions of a table in flight at any time (if it is being changed while others are viewing it).
For the sake of the is question let's assume that the table in question is large (contains a lot of data) which I expect is true. These two statements end up doing a common action - unlinking and freeing all the blocks in the table. The blocks is where all the data exists so you'd think that the speed of these two are the same and on idle systems they are close. Both automatically commit the results so the command doesn't complete until the work is done. In this idle system comparison I've seen DROP run faster but then you need to CREATE the table again so there is time needed to recreate the data structure of the table but this can be in a transaction block so do we need to include the COMMIT? The bottom line is that in the idle system these two approaches are quite close in runtime and when I last measured them out for a client the DROP approach was a bit faster. I would advise you to read on before making your decision.
However, in the real world Redshift clusters are rarely idle and in loaded cases these two statements can be quite different. DROP requires exclusive control over the table since it does not run inside of a transaction block. All other uses of the table must be closed (committed or rolled-back) before DROP can execute. So if you are performing this DROP/recreate procedure on a table others are using the DROP statement will be blocked until all these uses complete. This can take an in-determinant amount of time to happen. For ETL processing on "hidden" or "unpublished" tables the DROP/recreate method can work but you need to be really careful about what other sessions are accessing the table in question.
Truncate does run inside of a transaction but performs a commit upon completion. This means that it won't be blocked by others working with the table. It's just that one version of the table is full (for those who were looking at it before truncate ran) and one version is completely empty. The data structure of the table has versions for each session that has it open and each sees the blocks (or lack of blocks) that corresponds to their version. I suspect that it is managing these data structures and propagating these changes through the commit queue that slows TRUNCATE down slightly - bookkeeping. The upside for this bookkeeping is that TRUNCATE will not be blocked by other sessions reading the table.
The deciding factors on choosing between these approaches is often not performance, it is which one has the locking and coherency features that will work in your solution.
I'm in the process of evaluating some different data stores for a project and I have a strange but inflexible requirement to check the existence of a 1500 keys per query... Basically the only query I'll be running is of the form:
SELECT user_id, name, gender
WHERE user_id in (user1, user2, ..., user1500)
I will have around 3.5 billion rows in the table. One data store that has caught my eye is Spanner. I was wondering if querying the data in this way would be feasible or if I would run into performance issues due to the large number of items in my WHERE clause. I have only been able to test these queries on a small amount of data so far so I'm leaning more on what the theoretical performance hit might look like instead having the luxury to just "try and found out".
Also, are there other data stores that might work better for this read pattern? I expected to run no more than 80 queries per second. Also, the data will be bulk loaded on a weekly basis. The data is structured by nature but we don't use it in a relational way (i.e. no joins).
Anyways, sorry if this question is vague in any way. I'm happy to provide more detail if needed.
1500 keys should not be a problem if you use a bound array parameter to specify the keys:
SELECT user_id, name, gender
FROM table
WHERE user_id in UNNEST(#users)
https://cloud.google.com/spanner/docs/sql-best-practices#write_efficient_queries_for_range_key_lookup
I've included some links along with our approaches to other answers, which seem to be the most optimal on the web right now.
Our records need to be categorized (eg. "horror", "thriller", "tv"), and randomly accessible both in specific categories and across all/some categories. We generally need to access about 20 - 100 items at a time. We also have a smallish number of categories (less than 100).
We write to the database for uploading/removing content, although this is done in batches and does not need to be real time.
We have tried two different approaches, with two different data structures.
Approach 1
AWS DynamoDB - Pick a record/item randomly?
Help selecting nth record in query.
In short, using the category as a hash key, and a UUID as the sort key. Generate a random UUID, query Dynamo using greater than or less than, and limit to 1. This is even suggested by an AWS employee in the second link. (We've also tried increasing the limit to the number of items we need, but this increases the probability of the query failing the first time around).
Issues with this approach:
First query can fail if it is greater than/less than any of the UUIDs
Querying on any specific category will cause throttling at scale (Small number of partitions)
We've also considered adding a suffix to each category to artificially increase the number of partitions we have, as pointed out in the following link.
AWS Database Blog
Choosing the Right DynamoDB Partition Key
Approach 2
Amazon Web Services: How do we get random item from the dynamoDb's table?
Doing something similar to this, where we concatenate the category with a sequential number, and use this as the hash key. e.g. horror-000001.
By knowing the number of records in each category, we're able to perform random queries across our entire data set, while also avoiding hot partitions/keys.
Issues with this approach
We need a secondary data structure to manage the sequential counts across each category
Writing (especially deleting) is significantly more complex, although this doesn't need to happen in real time.
Conclusion
Both approaches solve our main use case of random queries on category/categories, but the cons they offer are really deterring us from using them. We're leaning more towards approach #1 using suffixes to solve the hot partitioning issue, although we would need the additional retry logic for failed queries.
Is there a better way of approaching this problem? Specifically looking for solutions capable of scaling well (No scan), without requiring extra resources be implemented. #1 fits the bill, but needing to manage suffixes and failed attempts really deters us from using it, especially when it is being called inside a lambda (billed for time used).
Thanks!
Follow Up
After more research and testing, my team has decided to move towards MySQL hosted on RDS for these tables. We learned that this is one of the few use cases were DynamoDB does not fit, and requires rewriting your use case to fit the DB (Bad).
We felt that the extra complexity required to integrate random sampling on DynamoDB wasn't worth it, and we were unable to come up with any comparable solutions. We are, however, sticking with DynamoDB for our tables that do not need random accessibility due to the price and response times.
For anyone wondering why we chose MySQL, it was largely due to the Nodejs library available, great online resources (which DynamoDB definitely lacks), easy integration via RDS with our Lambdas, and the option to migrate to Amazons Aurora database.
We also looked at PostgreSQL, but we weren't as happy with the client library or admin tools, and we believe that MySQL will suit our needs for these tables.
If anybody has anything else they'd like to add or a specific question please leave a comment or send me a message!
This was too long for a comment, and I guess it's pretty much a full fledged answer now.
Approach 2
I've found that my typical time to get a single item from dynamodb to a host in the same region is <10ms. As long as you're okay with at most 1-2 extra calls, you can quite easily implement approach 2.
If you use a keys only GSI where the category is your hash key and the primary key of the table is your range key, you can quickly find the largest numbered single item within a category.
When you add a new item, find the largest number for that category from the GSI and then write the new item to the table with sequence number n+1.
When you delete, find the item with the largest sequence number for that category from the GSI, overwrite the item you are deleting, and then delete the now duplicated item from its position at the highest sequence number.
To randomly get an item, query the GSI to find the highest numbered item in the category, and then randomly pick a number since you now know the valid range.
Approach 1
I'm not sure exactly what you mean when you say "without requiring extra resources to be implemented". If you're okay with using a managed resource (no dev work to implement), you can also make Approach 1 work by putting a DAX cluster in front of your dynamodb table. Then you can query to your heart's content without really worrying about hot partitions. (Though the caching layer means that new/deleted items won't be reflected right away.)
Related to this question, I'm looking for more a more specific answer. In an effort to keep this non-subjective, here is a full thought process for creating an activities table with a stuck point that can be finished with a quick example answer.
In an effort to better understand DynamoDB, I'm creating a personal website that contains an activity feed from a DynamoDB table. The goal is to evenly distribute partition keys while still being able to sort across all partition keys (I'm struggling with this part).
Different types of activities will include blog posts, projects, twitter post references, LinkedIn post references, etc. Using the activity type as a partition key would not be wise as my activity is highly weighted, mostly on the twitter side, hardly ever creating blog posts.
A unique activity id seems to be the best option for evenly distributing activities across DynamoDB partitions. However, this completely removes the ability to sort activities to start, as queries require a partition id to be known first. This is where a secondary global index (SGI) will be helpful. With this, a sort key will not be required on the primary partition key, but paired in an SGI.
This is part where I'm stuck. What do I base the SGI partition key on? At the moment I'm thinking of a single value "activity" for all activities with a sort key of "date", but that is a single partition for all entries. Will a single SGI partition key value limit performance in this project?
Note that this is a small scale project. However, I'm thinking about large scale projects while building this one, attempting to create the best DynamoDB table possible in regards to optimized partition distribution, while still keeping it flexible for sorting all table records.
Consider GSI (Global Secondary Index) same as Main Table indexes while designing your schema as they also get Read/Write provisioning limits and are subject to hot partition throttling as well which back pressures on main table in other words if your GSI gets throttled then your main table will start throttling requests.
Will a single SGI partition key value limit performance in this project?
Single partition for complete table is definitely misuse of DDB scalable capability.
The goal is to evenly distribute partition keys while still being able to sort across all partition keys (I'm struggling with this part).
You can sort across partitions using GSI but you will again need partition key for your GSI and if that partition key is not distributed enough then you get into problems I mentioned above.
DDB is powerful for put/get operations if modeled right and for fairly simple queries with some filters. In general, you will utilize your throughput more efficiently as the ratio of partition key values accessed to the total number of partition key values in a table grows.
For your specific need its not directly possible to get scalable solution from DDB but we still have few options
Option 1:
We can model the data such that it is fairly distributed for writes and will need extra work while reading it back, this pattern is also known as Randomizing Across Multiple Partition Key Values. Since you don't want to access specific item for given time this will work for us.
Idea is to create fixed set (say 1 to 100) and randomly pick a number from it to append to creation date (not timestamp) and have creation timestamps as sort key.
This will distribute your load across multiple random partitions but increases the read complexity as you will need to query all partitions and merge to get final sort view for that date.
Option 2:
Use multiple tables for hot and cold data as it is time series based data. For info read
http://docs.aws.amazon.com/amazondynamodb/latest/developerguide/GuidelinesForTables.html#GuidelinesForTables.TimeSeriesDataAccessPatterns
Option 3:
Scan? Not a good choice if we talk about scalability and when your data grows but for fairly small set of data it surely helps so mentioning it.
These are just an example not saying a good fit for your usecase.
So here is a thought process question for you: write down all your use-cases and access patterns. Figure out their importance which are fine with eventual consistency which are not and see if DDB is good fit for them at first place, don't be tempted to use DDB and then struggling with access pattern scalability.
Also read https://stackoverflow.com/a/38790120/962545 for more questions you must be asking yourself before restricting yourself for specific access pattern you want from DDB.
Don't forget to read best practices: http://docs.aws.amazon.com/amazondynamodb/latest/developerguide/GuidelinesForTables.html
I know there is a way to place the results of a query into a table; there is a way to copy a whole table into another table; and there is a way to list a table piecemeal (tabledata:list using startIndex, maxResults and pageToken).
However, what I want to do is go over an existing table with tabledata:list and output the results piecemeal into other tables. I want to use this as an efficient way to shard a table.
I cannot find a reference to such a functionality, or any workaround to it for that matter.
Important to realize: Tabledata.List API is not part of BQL (BigQuery SQL) but rather BigQuery API that you can use in client of your choice.
That said, the logic you outlined in your question can be implemented in many ways, below is an example (high level steps):
Calling Tabledata.List within the loop using pageToken for next iteration or for exiting loop.
In each iteration, process response from Tabledata.List, extract actual data and insert into destination table using streaming data with Tabledata.InsertAll API. You can also have inner loop to go thru rows extracted in given iteration and define which one to go to which table/shard.
This is very generic logic and particular implementation depends on client you use.
Hope this helps
For what you describe, I'd suggest you use the batch version of Cloud Dataflow:
https://cloud.google.com/dataflow/
Dataflow already supports BigQuery tables as sources and sinks, and will keep all data within Google's network. This approach also scales to arbitrarily large tables.
TableData.list-ing your entire table might work fine for small tables, but network overhead aside, it is definitely not recommended for anything of moderate size.