Data analysis often have two kind of results:
- Machine to machine result. Your output is being parsed by another system. No need for drawing here.
- Visualization. A human being will interpret a graphical result.
For visualization, you will output a json, which will be parsed and processed by your favorite chart/graph library, sometimes embedded in advanced dataviz solution, such as grafana, kibana, jupyter, zeppelin…
Sounds good. The limit of this approach is the number of datapoint. As long as you remain reasonable, your browser will handle the load of parsing a 100k of datapoint to plot lines + axis + …
Increasing the number of datapoints to visualize lead to a few problems:
- Bandwidth problem : 1 million of DOUBLES produce a 12MB json. Even using GZIP compression (default in Warp 10), you still have a 1MB file to transfer.
- Browser performance problem: handling such a json is not easy for graph libraries
- Approximation problem: even a high end UHD screen is only 3840 pixels wide. And graphic libs can sometimes have a weird behavior.
Display in your favorite display lib can quickly lead to:
The examples and figures below are based on a acceleration sensor time series. This acquisition is done at 50kHz and has 1’250’000 data points. Fetching data from a local Warp 10 instance take 1.7s, but plotting this acquisition in visual studio code (or any browser) is very painful. After 20s of patience, I have this image:
If I resize the window, interface freeze again during 20s. I need to do analytics to focus only on the red circles on the image above, so I need to be able to see or compare them at each step of my analysis. In the Warp 10™ toolbox, I can use BUCKETIZE to subsample, but Shannon quickly catch on me:
Also often used in the Warp 10™ toolbox, I can use LTTB with 1000 points. This algorithm intelligently keeps points that keep the global shape. It is very efficient to keep the anomaly or extrema. But again, there could be some surprise with a high frequency signal such as this one:
Drawing server side
WarpScript™ also includes a big subset of the Processing drawing library. It means you can draw an image in a WarpScript. I will try a very simple approach: plot a 1 pixel dot for every data point. Without any kind of optimization, I build a macro to NORMALIZE the curve, then plot it with the Ppoint function:
Click to see the WarpScript
<% SAVE 'context' STORE // save stack context 'gts' STORE $gts NORMALIZE 'normalizedgts' STORE //store the first and last tick $gts LASTTICK 'stopTS' STORE $gts FIRSTTICK 'startTS' STORE //compute the scale by pixel $width TODOUBLE $stopTS $startTS - / 'xscale' STORE $height TODOUBLE 'yscale' STORE $width $height '2D' PGraphics //create an image object 1.01 PstrokeWeight 0xffff9900 Pstroke //orange, 100% opacity 255 Pbackground //white background //store the raw list of ticks and values $normalizedgts TICKLIST 'ticks' STORE $normalizedgts VALUES 'values' STORE 0 $normalizedgts SIZE 1 - <% 'i' STORE //store the for loop index $ticks $i GET $startTS - $xscale * //x $height $values $i GET $yscale * - //y Ppoint %> FOR Pencode //push a base64 encoded png on the stack $context RESTORE //restore context %> 'drawplotgraph' STORE
Then I execute this macro on the same time series: In the Warp 10 JSON output, I have a nice string with a base64 encoded png image. Images on the stack are detected by VSCode WarpScript extension. Images are displayed in the VSCode “Images” tab.
Execution time is around 5s, and the result is really lightweight: only 14kB. For information, execution time of NORMALIZE takes around 100ms. The resulting image won’t freeze my IDE or my browser. But it is less readable.
How about drawing lines between each points? The CPU cost of drawing a line is higher than a simple one pixel dot. Code is a bit more complex, I keep the previous point on the stack and I use the Pline function.
Click to see the WarpScript
<% SAVE 'context' STORE // save stack context 'gts' STORE $gts NORMALIZE //[ SWAP bucketizer.last 0 0 1000 ] BUCKETIZE 0 GET 'normalizedgts' STORE $gts LASTTICK 'stopTS' STORE $gts FIRSTTICK 'startTS' STORE $width TODOUBLE $stopTS $startTS - / 'xscale' STORE $height TODOUBLE 'yscale' STORE $normalizedgts TICKLIST 'ticks' STORE $normalizedgts VALUES 'values' STORE //first point coordinates left on the stack $ticks 0 GET $startTS - $xscale * $height $values 0 GET $yscale * - $width $height '2D' PGraphics //create an image object 1.01 PstrokeWeight 0xff004EFF Pstroke //blue, 100% opacity 255 Pbackground //stack = pgrahics on top, previous y, previous x 0 $normalizedgts SIZE 1 - <% 'i' STORE $ticks $i GET $startTS - $xscale * DUP 5 ROLLD $height $values $i GET $yscale * - DUP 6 ROLLD 5 ROLL 5 ROLL //using stack instead of variables for performances. Pline %> FOR Pencode //push a base64 encoded png on the stack 3 ROLLD DROP DROP //remove previous x y of the stack $context RESTORE //restore context %> 'drawlinegraph' STORE
Image generation is done in 6.5s.
Compare results !
If I overlay the very first image with the one generated, both matches, but I also see the JS lib introduces some useless noise when the signal is near zero:
Why is there so much noise in the first seconds ? Because the dygraphs JS lib has a limit at high frequency. In the following WarpScript, I generate a GTS in two steps: a low frequency, and a high frequency. Amplitude is the same, +/-0.5.
Click to see the WarpScript
NOW 'startTS' STORE NEWGTS 'bmax' RENAME $startTS NaN NaN NaN 150.0 ADDVALUE NEWGTS 'bmin' RENAME $startTS NaN NaN NaN -150.0 ADDVALUE NEWGTS 'testGTS' RENAME $startTS $startTS 0.5 s + <% 5 ms + %> <% NaN NaN NaN <% RAND 0.5 < %> <% -0.5 %> <% 0.5 %> IFTE ADDVALUE %> FORSTEP $startTS 0.5 s + $startTS 1.2 s + <% 10 us + %> <% NaN NaN NaN <% RAND 0.5 < %> <% -0.5 %> <% 0.5 %> IFTE ADDVALUE %> FORSTEP
To avoid this bug, Warp 10 vizualisation component have a “date/timestamp” button. Switching to timestamp mode avoid this issue. Switching to steps also avoid the issue but has a cost : the lib must draw one extra line per datapoint.
Part 1 conclusion
- Drawing server side is a solution to visualize high frequency time series. Server rendering is faster than JS graphic libs.
- Warp 10 Processing functions provide you all the tooling to create nice images. From basic lines to bezier curves, with or without antialiasing.
- Do not blindly trust your favorite graphic library !
Electronics engineer, fond of computer science, embedded solution developer.