Context
I'm trying to find the best way to represent and aggregate a high-cardinality column in Redshift. The source is event-based and looks something like this:
user
timestamp
event_type
1
2021-01-01 12:00:00
foo
1
2021-01-01 15:00:00
bar
2
2021-01-01 16:00:00
foo
2
2021-01-01 19:00:00
foo
Where:
the number of users is very large
a single user can have very large numbers of events, but is unlikely to have many different event types
the number of different event_type values is very large, and constantly growing
I want to aggregate this data into a much smaller dataset with a single record (document) per user. These documents will then be exported. The aggregations of interest are things like:
Number of events
Most recent event time
But also:
Number of events for each event_type
It is this latter case that I am finding difficult.
Solutions I've considered
The simple "columnar-DB-friendy" approach to this problem would simply be to have an aggregate column for each event type:
user
nb_events
...
nb_foo
nb_bar
1
2
...
1
1
2
2
...
2
0
But I don't think this is an appropriate solution here, since the event_type field is dynamic and may have hundreds or thousands of values (and Redshift has a upper limit of 1600 columns). Moreover, there may be multiple types of aggregations on this event_type field (not just count).
A second approach would be to keep the data in its vertical form, where there is not one row per user but rather one row per (user, event_type). However, this really just postpones the issue - at some point the data still needs to be aggregated into a single record per user to achieve the target document structure, and the problem of column explosion still exists.
A much more natural (I think) representation of this data is as a sparse array/document/SUPER:
user
nb_events
...
count_by_event_type (SUPER)
1
2
...
{"foo": 1, "bar": 1}
2
2
...
{"foo": 2}
This also pretty much exactly matches the intended SUPER use case described by the AWS docs:
When you need to store a relatively small set of key-value pairs, you might save space by storing the data in JSON format. Because JSON strings can be stored in a single column, using JSON might be more efficient than storing your data in tabular format. For example, suppose you have a sparse table, where you need to have many columns to fully represent all possible attributes, but most of the column values are NULL for any given row or any given column. By using JSON for storage, you might be able to store the data for a row in key:value pairs in a single JSON string and eliminate the sparsely-populated table columns.
So this is the approach I've been trying to implement. But I haven't quite been able to achieve what I'm hoping to, mostly due to difficulties populating and aggregating the SUPER column. These are described below:
Questions
Q1:
How can I insert into this kind of SUPER column from another SELECT query? All Redshift docs only really discuss SUPER columns in the context of initial data load (e.g. by using json_parse), but never discuss the case where this data is generated from another Redshift query. I understand that this is because the preferred approach is to load SUPER data but convert it to columnar data as soon as possible.
Q2:
How can I re-aggregate this kind of SUPER column, while retaining the SUPER structure? Until now, I've discussed a simplified example which only aggregates by user. In reality, there are other dimensions of aggregation, and some analyses of this table will need to re-aggregate the values shown in the table above. By analogy, the desired output might look something like (aggregating over all users):
nb_events
...
count_by_event_type (SUPER)
4
...
{"foo": 3, "bar": 1}
I can get close to achieving this re-aggregation with a query like (where the listagg of key-value string pairs is a stand-in for the SUPER type construction that I don't know how to do):
select
sum(nb_events) nb_events,
(
select listagg(s)
from (
select
k::text || ':' || sum(v)::text as s
from my_aggregated_table inner_query,
unpivot inner_query.count_by_event_type as v at k
group by k
) a
) count_by_event_type
from my_aggregated_table outer_query
But Redshift doesn't support this kind of correlated query:
[0A000] ERROR: This type of correlated subquery pattern is not supported yet
Q3:
Are there any alternative approaches to consider? Normally I'd handle this kind of problem with Spark, which I find much more flexible for these kinds of problems. But if possible it would be great to stick with Redshift, since that's where the source data is.
Related
I have a system with a large number of tables that contain historical data. Each table has a ts_from and ts_to column which are of type timestamptz. These represent the time period in which the data for a particular row was valid.
These columns are indexed.
If I want to query all rows that were valid at a particular timestamp, it is trivial to write the ts_from <= #at_timestamp AND ts_to >= #at_timestamp WHERE clause to utilitise the index.
However, I wanted to create a function called Temporal.at which would take the #at_timestamp column and the ts_from / ts_to columns and do this by hiding the complexity of the comparison from the query that uses it. You might think this is trivial, but I would also like to extend the concept to create a function called Temporal.between which would take a #from_timestamp and #to_timestamp and select all rows that were valid between those two periods. That function would not be trivial, as one would have to check where rows partially overlap the period rather than always being fully enclosed by it.
The issue is this: I have written these functions but they do not cause the index to be used. The query performance is woefully slow on the history tables, some of which have hundreds of millions of rows.
The questions therefore are:
a) Is there a way to write these functions so that we can be sure the indexes will be used?
b) Am I going about this completely the wrong way and is there a better way to proceed?
This is complicated if you model ts_from and ts_to as two different timestamp columns. Instead, you should use a range type: tstzrange. Then everything will become simple:
for containment in an interval, use #at_timestamp <# from_to
for interval overlap, use tstzinterval(#from_timestamp, #to_timestamp) && from_to
Both queries can be supported by a GiST index on the range column.
Let me give a summary of what I've been attempting to do and the efficiency issues I've been running into:
Essentially I want my users to be able to select parameters to filter data from my database, then I want to pass relevant data which passes those filters from the controller.
However, these filters query on data from multiple different tables (that is, about 5-6 different tables), some of which are quite large (as in 100k+ rows). These tables are all related to what I want to show, e.g. Here is a bond that meets so and so criteria, which is issued by so and so issuer, which must meet these criteria, and so on.
From an end result, I only really need about 100 rows after querying based on the parameters given by the user, but it feels like I need to look at everything in every table because I dont know how strict the filters will be beforehand. e.g. With a starting universe of 100k sets of data, passing filter f1,f2 of Table 1 might leave 90k, but after passing through filter f3 of table 2, f4,f5,f6 of table 3, and so ..., we might end up with 100 or less sets of data that pass these parameters because the last filters checked might be quite strict.
How can I go about querying through these multiple different tables efficiently?
Doing a join between them seems like it'd yield some time complexity of |T_1||T_2||T_3||T_4||T_5||T_6| where T_i is the "size" of table i.
On the other hand, just looking through the other tables based off the ids of the ones that pass the previous filter (as in, id 5,7,8 pass filters in T_1, which of those ids then pass filters in T_2, then which of those pass filters in T_3 and so on) looks like it might(?) have time complexity of |T_1| + |T_2| + ... + |T_6|.
I'm relatively new to Ruby on Rails, so im not entirely sure all of the tools at my disposal that could help with optimizing this, but at the same time I'm not entirely sure how to best approach this algorithmically.
I have this kind of data:
I need to transpose this data into something like this using Talend:
Help would be much appreciated.
dbh's suggestion should work indeed, but I did not try it.
However, I have another solution which doesn't require to change input format and is not too complicated to implement. Indeed the job has only 2 transformation components (tDenormalize and tMap).
The job looks like the following:
Explanation :
Your input is read from a CSV file (could be a database or any other kind of input)
tDenormalize component will Denormalize your column value (column 2), based on value on id column (column 1), separating fields with a specific delimiter (";" in my case), resulting as shown in 2 rows.
tMap : split the aggregated column into multiple columns, by using java's String.split() method and spreading the resulting array into multiple columns. The tMap should like like this:
Since Talend doesn't accept to store Array objects, make sure to store the splitted String in Object format. Then, cast that object into Array on the right side of the Map.
That approach should give you the expected result.
IMPORTANT:
tNormalize might shuffle the rows, meaning for bigger input, you might encounter unsorted output. Make sure to sort it if needed or use tDenormalizeSortedRow instead.
tNormalize is similar to an aggregation component meaning it scans the whole input before processing, which results into possible performance issues with particularly big inputs (tens of millions of records).
Your input is probably wrong (you have 5 entries with 1 as id, and 6 entries with 2 as id). 6 columns are expected meaning you should always have 6 lines per id. If not, then you should implement dbh's solution, and you probably HAVE TO add a column with a key.
You can use Talend's tPivotToColumnsDelimited component to achieve this. You will most likely need an additional column in your data to represent the field name.
Like "Identifier, field name, value "
Then you can use this component to pivot the data and write a file as output. If you need to process the data further, read the resulting file with tFileInoutDelimited .
See docs and an example at
https://help.talend.com/display/TalendOpenStudioComponentsReferenceGuide521EN/13.43+tPivotToColumnsDelimited
I am new to postgres and am experimenting with the hstore extension.Looking for some guidance. I need to support basic reporting on timeseries data for various products that we sell. I have a large amount data in the format "Timestamp, Value" for each product. This data is available in a csv fle for each product.
I am thinking of using hstore to store this data in the key value format. Assuming that all the timeseries data for a single product can be stored in a single hstore object. I need to be able to query this data by specific times, say what was the value of a product at a given time? Also need to run simple queries like retrieving the times where the product costed more than $100.
I'm planning to have a table with a product id column and an hstore column. But I am not very clear on how to make this work:
The hstore column needs to be loaded from thousands of timestamp,value records that exist in a csv. The hstore should be appended whenever we get a new csv.
The table needs to store the productId and corresponding Timeseries data.
Can you please advise if using hstore would be helpful ? If yes then how can I load data from csv as explained above. Also, if there could be any impact on the performance on inserts/updates in the hstore, as data grows please share your experiences.
I do think you should start with a simple, normalised schema first, especially since you are new to PostgreSQL. Something like:
CREATE TABLE product_data
(
product TEXT, -- I'm making an assumption about the types of your columns
time TIMESTAMP,
value DOUBLE PRECISION,
PRIMARY KEY (product, time);
);
I would definitely keep hstore and similar options in mind, if and when your data becomes large enough that efficiency is more important and simplicity. But note that all options have an efficiency tradeoff.
Do you know how much data you're going to support? Number of products, number of distinct timestamps for each product?
What other queries do you want to run? A query for the times where a single product cost more than $100 would benefit from an index on (product, value), if the product has many distinct timestamps.
Other options
hstore is most useful if you want to store a table set of arbitrary key-value pairs in a row. You could use it here, with a row for each product, and each distinct timestamp for that product being a key in the product's table. The downsides are that keys and values in hstore are text, whereas your keys are timestamps, and your values are numbers of some kind. So there will be a certain reduction in type checking, and a certain increase in type casting cost required. Another possible downside is that some queries on the hstore might not use indexes very efficiently. The above table can use simple btree indexes for range queries (say you want to pull out the values between two dates for a product). But hstore indexes are much more limited; you can use a gist or gin index on an hstore column to find all the rows that feature a certain key.
Another option (which I've played with and use experimentally for some of my databases) is arrays. Basically, each product will have an array of values, and each timestamp maps to an index in the array. This is easy if the timestamps are perfectly regular. For example, if all your products had a value every hour for every day, you could use a table like this:
CREATE TABLE product_data
(
product TEXT,
day DATE,
values DOUBLE PRECISION[], -- An array from 0 to 23.
PRIMARY KEY (product, day);
);
You can construct views and indexes to make querying this table moderate easy. (I wrote a blog post on this technique at http://ejrh.wordpress.com/2011/03/20/vector-denormalisation-in-postgresql/.)
But my advice is still: start with a simple table, then explore ways to improve efficiency when you know you're going to need them.
I have a table that has around 40 columns. The only difference in the columns names is that the last 20 all start with "B" before the column name. This table is used for comparing. In other words, compare the data in the first 20 columns to the data in the last 20 columns.
I know this is very bad design, so how should this table be redesigned, so that there are only 20 columns, yet we can still compare the data?
EDIT: if it helps, we also use this data to find a matched cohort
Also note that performance is of main concern here. By duplicating the columns the getting of data is extremely fast.
Thanks!
Two possible architectures and a query tip.
1) Build your table with a "Type" column, and use that to flag "primary" vs. "alternate". In your case, "A" vs. "B" might be appropriate.
2) Build a vertical partition, two identical tables (for primary and alternate data), that share a common primary key. (If Id = 42 is in one table, it must be in the other--unless "alternate" data is optional, in which case don't populate the second table.) Also optionally, have a third table that tracks all possible primary keys, along with any data that is known to always be common to both tables.
Tip: Read up on SELECT...EXCEPT and SELECT...INTERSECT. They run disturbingly quickly, and are idea for comparing all columns and rows between two datasets for differences (except) and matches (intersect). You can use this fairly easily with either of the two structures, and it would work with your existing code as well (though it might be fussier to write the query).