The not so obvious difference in the Tgres approach, and my blogs might not be doing a great job of explaining it, is that some time in Feb 2017 I significantly revamped the storage approach to what I dubbed "vertical" storage whereby a timeslot stores an array of points in which every array element represents an element of a(nother) series.
So it went from:
series1, array[val1, val2, val3 ...] --> time direction
series2, array[val1, val2, val3 ...]
...
to
slot1, array[series1_val1, series2_val1, ...] |
slot2, array[series1_val2, series2_val2, ...] |
^ time dir
With this structure you can write a single row and insert data points for n series (where n is array length) in a single row insert.
Thus, if I have 10,000 series, and my arrays are 1000-long, I can insert a data point for each of the 10K series in only 10 row inserts. This only works if the data points for all series for the same slot arrive at approximately same time, which in a monitoring-like scenario they usually do, but in other situations might not be the case.
The flip side of this approach is that querying the data then becomes less efficient because to read one data point of a series you end up reading an array-length of data points you might not care about for this particular query.
Also, tgres takes the round-robin approach, versus the timed partition approach and that's completely apples and oranges when it comes to performance. The round-robin approach also works only if the data points are evenly spaced (or transformed to be evenly spaced on the fly, which is what tgres Go code does), and again, it's hard to judge whether that's fundamentally "good" or "bad".
I can see how readers of this thread my be looking for which technique is faster, but it's just not that simple, and very much depends on the what the actual requirements are. The round-robin versus timed partition is also not mutually exclusive, you can combine the two, which may or may not be faster, not sure, the devil is in the details.
> But DBs and FSs operate on pages of data and not individual records, so you will be reading that row anyway, and likely much more.
Yes, when dealing with database performance understanding this goes with the territory.
The "game" here is to organize data in such way that the stuff you read inadvertently is something that you will need eventually (as in in a few microseconds). This is where things like CLUSTER and BRIN indexes become important, and this is also why partitioning is a win.
So it went from:
to With this structure you can write a single row and insert data points for n series (where n is array length) in a single row insert.Thus, if I have 10,000 series, and my arrays are 1000-long, I can insert a data point for each of the 10K series in only 10 row inserts. This only works if the data points for all series for the same slot arrive at approximately same time, which in a monitoring-like scenario they usually do, but in other situations might not be the case.
The flip side of this approach is that querying the data then becomes less efficient because to read one data point of a series you end up reading an array-length of data points you might not care about for this particular query.
Also, tgres takes the round-robin approach, versus the timed partition approach and that's completely apples and oranges when it comes to performance. The round-robin approach also works only if the data points are evenly spaced (or transformed to be evenly spaced on the fly, which is what tgres Go code does), and again, it's hard to judge whether that's fundamentally "good" or "bad".
I can see how readers of this thread my be looking for which technique is faster, but it's just not that simple, and very much depends on the what the actual requirements are. The round-robin versus timed partition is also not mutually exclusive, you can combine the two, which may or may not be faster, not sure, the devil is in the details.