How does Amazon Redshift reconstruct a row from columnar storage? - amazon-web-services

Amazon describes columnar storage like this:
So I guess this means in what PostgreSQL would call the "heap", blocks contain all the values for one column, then the next column, and so on.
Say I want to query for all people in their 30's, and I want to know their names. So columnar storage means less IO is required to read just the age of every row and find those that are 30-something, because all the other columns don't need to be read. Also maybe some efficient compression can be applied. That's neat, I guess.
Then what? This data structure alone doesn't explain how anything useful can happen after that. After determining what records are 30-something, how are the associated names found? What data structure is used? What are its performance characteristics?

If the Age column is the Sort Key, then the rows in the table will be stored in order of Age. This is great, because each 1MB storage block on disk keeps data for only one column, and it keeps note of the minimum and maximum values within the block.
Thus, searching for the rows that contain an Age of 30 means that Redshift can "skip over" blocks that do not contain Age=30. Since reading from disk is the slowest part of a database, this means it can operate much faster.
Once it has found the blocks that potentially contain Age=30, it reads those blocks from disk. Blocks are compressed, so they might contain much more data than the 1MB on disk. This means many rows can be read with fewer disk accesses.
Once those blocks are decompressed into memory, it finds the rows with Age=30 and then loads the corresponding blocks for the Name column. The compression ratio would be different for the Name column since it is text and is not sorted, so this might result in loading more blocks from disk for Name than for Age.
Redshift then assembles the data from Name and Age for the desired rows and performs any remaining operations.
These operations are also parallelized across multiple nodes based on the Distribution Key, which distributed data based on a given column (or replicates it between nodes for often-used tables). Data is typically distributed based upon a column that is frequently used in JOIN statements so that similar data is co-located on the same node. Each node returns its data to the Leader Node, which combines the data and provides the final results.
Bottom line: Minimise the amount of data read from disk and parallelize operations on separate nodes.

AFAIK every value in the columnar storage has an ID pointer (similar to CTID you mentioned), and to get the select results Redshift needs to find and combine the values with the same ID pointer for each column that's selected from the raw data. If memory allows it's stored in memory, unless it's spilling to disk. This process is called materialization (don't confuse with materialized view materialization). In your case there are 2 technically possible scenarios:
materialize all Age/Name pairs, then filter by Age=30, and output the result
filter Age column by Age=30, get IDs, get Name values with corresponding IDs, materialize pairs and output
I guess in this case #2 is what happens because materialization is more expensive than filtering. However, there is a plenty of scenarios where this is much less obvious (with complex queries and aggregations). It is the responsibility of the query optimizer to decide what's better. #1 is still better than the row oriented because it would still read just 2 columns.

Related

Aws RedShift sampling

For a data quality check I need to collect data in a specific interval.
Some tables are huge in size.
Is there any hack to do this without affecting the performance?
Like select 100 rows randomly.
How random do you need? The classic way to do this is with "WHERE RANDOM() < .001". If you need it to give you a repeatable "random" set then you can add a seed. The issue is that your tables are huge and this means reading (scanning) every row from disk just to throw most of them away and since table scan can take a significant time this isn't what you want to do.
So you may want to take advantage of Redshift "limited table scan" capabilities as part of your "random" sampling. (The fastest data to read from disk is the data you don't read from disk.) The issue here is that this solution will depend on your table sort keys and ordering which will push the solution into even "more pseudo" random territory (less of a true random sampling). In many cases this isn't a big deal but if the statistics really matter then this may not work for you.
This is done by sampling "blocks", not rows, based on the sort key(s). This sampling of blocks can be done randomly and each block of data will represent about 250K rows (based on sort key data type, compression etc. and COULD range anywhere from <100K rows to 2M rows). Doing this process will take a little inspection of STV_BLOCKLIST. The storage quanta for Redshift is the 1MB block and each and every block's metadata in the system can be referenced in STV_BLOCKLIST. This system table contains min and max values for each block. First find all the blocks for the sort key for the table in question. Next pick a random sample of these blocks (and if you are still dealing with a lot of data make sure that this sampling picks an even number from across all the slices to avoid execution skew).
Now the trick is to translate these min a max metadata values into a WHERE clause the performs the desired sampling. These min and max values are BIGINTs and are hashed from the data in the sort key column. This hash is data type dependent. If the data type is BIGINT then the has is quite simple - if the data type is timestamp then it is a bit more complex. But the ordering will be preserved across the hashing function for the data type involved. Reverse engineering this hash isn't hard - just perform a few experiments - but I can help if you tell me the type involved as I've done this for just about every data type at this point.
You can even do a random sampling rows on top of this random sampling of blocks. Or if you want you can just pick some narrow ranges of the sort key value and then randomly sample row and avoid all this reverse engineering business. The idea is to use Redshift "reduced scan" capability to greatly reduce the amount of data read from disk. To do this you need to be metadata aware in your choice of sampling windows which often means a sort key where clause. This is all about understanding how the database engine works and using its capabilities to your advantage.
I understand that this answer is based on some unstated information so please reach out in a comment if something isn't clear.

Redshift table sizes & flavours of

Confused by the term 'table size' in Redshift.
We have :
svv_table_info.size
"Size of table in 1MB blocks"
svv_table_info.pct_used
"Percent of available space used"
... so I assume that a lot of the 'size' is empty space due to sort keys etc
Then we have this..
https://aws.amazon.com/premiumsupport/knowledge-center/redshift-cluster-storage-space/
.. which uses the term 'minimum' table size.
But nowhere can I find an explanation of what they means in the real world ? Is this a theoretical minimum if optimally configured ?
Ultimately I need to find out the basic size of original tangible data without any overheads.
Then yes, how much disc space is it actually costing to store it in Redshift.
So if I took 1TB out of our on-prem database and shoved it into Redshift, I'd be looking to see something like 1TB (data) & 1.2TB (data + Redshift overheads).
Hope someone can help clarify 🤞
Redshift stores data in 1MB blocks and blocks are associated with a slice and a column. So if I have 2 slices in my cluster and a table with 4 columns (plus the 3 system columns to make 7) distributed as EVEN containing at least 2 rows, then my table will minimally take up 2 X 7 X 1MB of space (14MB on disk). This is all that article is saying.
Now if I insert 2 additional rows into this table, Redshift will makes new blocks for this data. So now my 4 rows of data take up 28MB of space. However, if I Vacuum the table the wasted space will be reclaimed and the table size will come back down to 14MB. (yes this is a bit of an oversimplification but trying to get the concepts across)
As a rule of thumb a single 1MB block will typically hold between 100,000 rows and 2,000,000 rows of compressed data. (yes this depends on the data not being monster varchars) So for our table above I can keep adding rows (and vacuuming) without increasing the table size on disk until I get a few hundred thousand rows (per slice) in the table. Redshift is very efficient at storing large chunks of data but very inefficient at storing small ones.
What Redshift knows about your data size is how many blocks it takes on disk (across all the nodes, slices, and columns). How big your data would be if it was stored differently (not in blocks, compressed or uncompressed) is not data that is tracked. As John noted, for big tables, Redshift stores data more efficiently than most other database (when compression is used).
You cannot translate from an existing database size to the size of a table in Redshift. This is because:
Columns are stored separately
Minimum block size is 1MB
Data in Redshift is compressed, so it can take considerably less space depending on the type of data and the compression type chosen
Given compression, your data is likely to be smaller in Redshift than an original (uncompressed) data source. However, you can't really calculate that in advance unless you have transferred similar data in the past and apply a similar ratio.

Is it worth dropping stale columns on a large data set?

I have a relatively large data set in a table with about 60 columns, of which about 20 have gone stale. I've found a few posts on dropping multiple columns and the performance of DROP COLUMN, but nothing on whether or not dropping a bunch of columns would result in a noticeable performance increase.
Any insight as to whether or not something like this could a perceptible impact?
Dropping one or more columns can be done in a single statement and is very fast. All it needs is a short ACCESS EXCLUSIVE lock on the table, so long running queries would block it.
The table is not rewritten during this operation, and it will not shrink. Subsequent rewrites (with VACUUM (FULL) or similar) will get rid of the column data.

Why Amazon Redshift UNLOAD performance is much better for fresh data?

I wonder why unloading from a big table (>100 bln rows) when selecting by a column, which is NOT a sort key or a part of sort key, is immensely faster for newly added data. How Redshift understands that it is time to stop sequential scan in the second scenario?
Time the query spent executing. 39m 37.02s:
UNLOAD ('SELECT * FROM production.some_table WHERE daytime BETWEEN
\\'2017-01-15\\' AND \\'2017-01-16\\'') TO ...
vs.
Time the query spent executing. 23.01s :
UNLOAD ('SELECT * FROM production.some_table WHERE daytime BETWEEN
\\'2017-06-24\\' AND \\'2017-06-25\\'') TO ...
Thanks!
Amazon Redshift uses zone maps to identify the minimum and maximum value stored in each 1MB block on disk. Each block only stores data related to a single column (eg daytime).
If the SORTKEY is not set to daytime, then the data is unsorted and any particular date could appear in many different blocks. If SORTKEY is used, then a particular date will only appear in a minimum number of blocks.
Your second query possibly executes faster, even without a SORTKEY, because you are querying data that was probably added recently and is therefore all stored together in just a few blocks. The historical data might be spread in many blocks because a VACUUM probably reordered the data based upon the correct SORTKEY. In fact, if you did a VACUUM now, you might find that your second query becomes slower.

Storing Time Series in AWS DynamoDb

I would like to store 1M+ different time series in Amazon's DynamoDb database. Each time series will have about 50K data points. A data point is comprised of a timestamp and a value.
The application will add new data points to time series frequently (all the time) and will retrieve (usually the whole time series) time series from time to time, for analytics.
How should I structure the database? Should I create a separate table for each timeseries? Or should I put all data points in one table?
Assuming your data is immutable and given the size, you may want to consider Amazon Redshift; it's written for petabyte-sized reporting solutions.
In Dynamo, I can think of a few viable designs. In the first, you could use one table, with a compound hash/range key (both strings). The hash key would be the time series name, the range key would be the timestamp as an ISO8601 string (which has the pleasant property that alphabetical ordering is also chronological ordering), and there would be an extra attribute on each item; a 'value'. This gives you the abilty to select everything from a time series (Query on hashKey equality) and a subset of a time series (Query on hashKey equality and rangeKey BETWEEN clause). However, your main problem is the "hotspot" problem: internally, Dynamo will partition your data by hashKey, and will disperse your ProvisionedReadCapacity over all your partitions. So you may have 1000 KB of reads a second, but if you have 100 partitions, then you have only 10 KB a second for each partition, and reading all data from a single time series (single hashKey) will only hit one partition. So you may think your 1000 KB of reads gives you 1 MB a second, but if you have 10 MB stored it might take you much longer to read it, as your single partition will throttle you much more heavily.
On the upside, DynamoDB has an extremely high but costly upper-bound on scaling; if you wanted you could pay for 100,000 Read Capacity units, and have sub-second response times on all of that data.
Another theoretical design would be to store every time series in a separate table, but I don't think DynamoDB is meant to scale to millions of tables, so this is probably a no-go.
You could try and spread out your time series across 10 tables where "highly read" data goes in table 1, "almost never read data" in table 10, and all other data somewhere in between. This would let you "game" the provisioned throughput / partition throttling rules, but at a high degree of complexity in your design. Overall, it's probably not worth it; where do you new time series? How do you remember where they all are? How do you move a time series?
I think DynamoDB supports some internal "bursting" on these kinds of reads from my own experience, and it's possible my numbers are off, and you will get adequete performance. However my verdict is to look into Redshift.
How about dripping each time series into JSON or similar and store in S3. At most you'd need a lookup from somewhere like Dynamo.
You still may need redshift to process your inputs.