This question is close, but doesn't quite help me with a similar issue as I am using a single data set and no related time series.
I am using AWS Forecast with a single time series dataset (no related data, just the main DS). It is a daily data set with about 10 years of data ranging from 2010-2020.
I have 3572 data points in the original data set; I manually filled missing data to ensure there were no missing days in the date range for a total of 3739 data points. I lopped off everything in 2020 to create a validation dataset and then configured the predictor for a 180 day Forecast. I keep getting the following error:
Unable to evaluate this dataset because there is missing data in the evaluation window for all items. Ensure that there is complete data for at least one item in the evaluation window starting from 2019-03-07T00:00:00 up to 2020-01-01T00:00.
There is definitely no missing data, I've double and triple checked the date range and data fill and every day between start and end dates has a data point. I also tried adding a data point for 1/1/2020 (it ended at 12/31/2019) and I continue to get this error. I can't figure out what it's asking me for, except that maybe I'm missing something in my math about the forecast Horizon and Backtest window offset?
Dataset example:
Brief model parameters (can share more if I'm missing something pertinent):
Total data points in training data: 3479
forecastHorizon = 180
create_predictor_response=forecast.create_predictor(PredictorName=predictorName,
ForecastHorizon=forecastHorizon,
PerformAutoML= True,
PerformHPO=False,
EvaluationParameters= {"NumberOfBacktestWindows": 1,
"BackTestWindowOffset": 180},
InputDataConfig= {"DatasetGroupArn": datasetGroupArn},
FeaturizationConfig= {"ForecastFrequency": 'D'
I noticed you don't have entry for 6/24/10 (this american date format is the worst btw)
I faced a similar problem when leaving out days (assuming you're modelling in daily frequency) just like that and having the Forecast automatic filling of gaps to nan values (as opposed to zero which is the default). I suggest you:
pre-fill literally every date within the range of training data (and of forecast window, if using related data)
choose zero as the option for automatically filling of missing values. I think mean or any other float value would also work for that matter
let me know if that works! I am also using Forecast and it's good to keep track of possible problems and solutions
Related
We are using amCharts 4 to show trend logs, and sometimes we end up with a lot of data that has to go into the chart. We'd like to know what the maximum number of data points that a chart can handle so we know how much data to aggregate (to reduce the data point count) before sending it into the package. To show the most accurate representation of the data as possible, we don't want to aggregate more aggressively than we have to. Our charts are x/y charts with value vs. date/time for up to 8 series.
In one case, we have a data set with well in excess of 600,000 data points in 8 series, and loading this into the chart, even in batches (i.e., loading one batch in, then adding the remaining batches to it in turn), will cause the charting package to run out of memory. In the case cited here, during our test, the charting package ran out of memory on the third batch, where the total of the 3 batches exceeded 600,000 data points, preventing further batches from being loaded in. For large sites that use our product, it is quite common to have that much data that the user wants to see in a chart if they want to see 6 months or a year's worth of data; so it's important that we be able to show some kind of representation of all that data, which is where aggregation comes in.
I have an event dataset in QuickSight, where each record has a timestamp field as following:
last_day_record_ts |
-------------------|
2020-01-19 05:46:55|
2020-01-20 05:55:37|
2020-01-21 06:00:12|
2020-01-22 06:12:57|
2020-01-23 06:02:15|
2020-01-24 06:15:35|
2020-01-25 06:20:05|
2020-01-26 05:55:48|
I want to build a visualization of time values over days as a line chart as following:
However, I find it difficult to get this in AWS QuickSight. Any ideas?
Instead of desired result QuickSight persistently gives just aggregated record values (i.e 1 for each day) but not the time values itself...
UPDATE. The workaround I found for now - to add calculated fields to the Data Set in order to get numeric values instead of timestamp ones.
Calculated fields:
day_midnight | truncDate('DD',{last_day_record_ts})
time_diff_in_hours_dec | abs(dateDiff({last_day_record_ts},{day_midnight},"MI")) / 60
time_diff_in_hours_int | decimalToInt({time_diff_in_hours_dec})
time_diff_in_min | ({time_diff_in_hours_dec} - {time_diff_in_hours_int}) * 60
The only problem I still cannot solve - to get Y axis labels in HH:MM format as in green rectangle. For now, it's numeric decimals...
Unfortunately, (after many attempts of my own) this type of visual does not appear to be possible in Quicksight at the time of writing.
Quicksight has many nice features, but it's still missing some (very basic imo) things that make it limiting for anyone working with data that is outside the expected use-cases.
While training my predictor I came across this error and I got stuck how to fix it.
I have two data-series, a "Target time-series data" with 9234 rows and a single "item_id" and a second one that is "Related time-series data" with the same number of rows as I only have a single id.
I'm setting de data with a window of 180 days, what is exactly the difference between the second and the first number that has appeared on the error, 9414 - 9234 = 180.
We were unable to train your predictor.
Please ensure there are no missing values for any items in the related time series, All items need data until 2020-03-15 00:00:00.0. For example, following items have missing data: item: brl only has 9234/9414 required datapoints starting 1994-06-07 00:00:00.0, please refer to documentation for additional details.
Once my data don't have missing data and it's on a daily basis why is it returning this error?
My data starts on 1994-06-07 and ends on 2019-09-17. Why should I have 9414 data points rather than 9234?
Should I take out 180 days in my "Target time-series data"?
The future values of the related time-series data must be known.
Example of a good related-time series: You know past and future days in which marketing has or will send email newsletters promoting the product you're forecasting. You can use this data as a related-time series.
Example of a bad related-time series: You notice that Google searches for your brand correlated with the sale of your product. As a result you want to use it as a related-time series. Since you don't know how many searches will occur in the future, so you can't use this as a related time series.
In you case, You have TARGET_TIME_SERIES data for 9414 days and you want to predict demand for the next 180 days. That means your RELATED_TIME_SERIES data should be 9594 days.
Edit: I have not tested this with amazon's forecasting product. I'm basing my answer on working with Facebook Prophet (which is one of the models amazon forcast uses). Please let me know if my solution worked.
Reading upon many Kimball design tips regarding fact tables (transaction, accumulating, periodic) etc. I'm still vague what should I do with my case of updating a fact table which I believe is not that uncommon. To the case.
We're processing complaints from clients, and we want to be able to reflect current status of complaint in the Data Warehouse. Our complaints have a workflow of statuses they go through, different assignees that deal with them on time, but for our analysis this is irrelevant as of now. We would like to review what the current situation on complaint is.
To my understanding the grain of the fact table would be single complaint, with columns (irrelevant for this question whether it should be junk dimension, degenerate etc) such as:
Complaint Number
Current Status
Current Status Date
Current Assignee
Type of complaint
As far as I understand, since we don't want to view the process history, but instead see what the current status of the process is, storing multiple rows for each complaint representing it's state is an overkill, so instead we store only one row per complaint and update it.
Now, is my reasoning correct to do that? In above case, complaint number and type of complaint store values that don't change, while "Current" columns do and we need to update the row, so we could implement Change Data Capture mechanism (just like we do for dimensions right now) to compare incoming rows from source system for this fact with currently stored fact rows to improve time cost of such operation.
It honestly looks like a Dimension table with mixed SCD Type 0&1 for me, but it stores facts of receiving complaints.
SO Post for reference: Fact table with information that is regularly updatable in source system
Edit
I'm aware that I could use accumulating fact table with time stamps which is somewhat SCD Type 2 alike but the end user doesn't really care about the history of the process. There are more facts involved in the analysis later on, so separating this need from data warehouse doesn't really work in this case.
I’ve encountered similar use cases in the past, where an accumulating snapshot would be the default solution.
However, the accumulating snapshot doesn’t allow processes with varying length. I’ve designed a different pattern, when 2 rows are added for each event: if an object goes from state A to state B you first insert a row with state A and quantity -1, then a new one with state B and quantity +1.
The end result allows:
- no updates necessary, only inserts;
- map-reduce friendly;
- arbitrary length processes;
- counting how many of each in each state at any point in time (with the help of a periodic snapshot for performance reasons);
- how many entered or left any state at any point in time.;
- calculate time in each state and age overall.
Details in 5 blog posts here (with implementation in Pentaho Data Integration):
http://ubiquis.co.uk/dwh/status-change-fact-table-part-1-the-problem/
I am having real problems getting the AWS IoT Analytics Delta Window (docs) to work.
I am trying to set it up so that every day a query is run to get the last 1 hour of data only. According to the docs the schedule feature can be used to run the query using a cron expression (in my case every hour) and the delta window should restrict my query to only include records that are in the specified time window (in my case the last hour).
The SQL query I am running is simply SELECT * FROM dev_iot_analytics_datastore and if I don't include any delta window I get the records as expected. Unfortunately when I include a delta expression I get nothing (ever). I left the data accumulating for about 10 days now so there are a couple of million records in the database. Given that I was unsure what the optimal format would be I have included the following temporal fields in the entries:
datetime : 2019-05-15T01:29:26.509
(A string formatted using ISO Local Date Time)
timestamp_sec : 1557883766
(A unix epoch expressed in seconds)
timestamp_milli : 1557883766509
(A unix epoch expressed in milliseconds)
There is also a value automatically added by AWS called __dt which is a uses the same format as my datetime except it seems to be accurate to within 1 day. i.e. All values entered within a given day have the same value (e.g. 2019-05-15 00:00:00.00)
I have tried a range of expressions (including the suggested AWS expression) from both standard SQL and Presto as I'm not sure which one is being used for this query. I know they use a subset of Presto for the analytics so it makes sense that they would use it for the delta but the docs simply say '... any valid SQL expression'.
Expressions I have tried so far with no luck:
from_unixtime(timestamp_sec)
from_unixtime(timestamp_milli)
cast(from_unixtime(unixtime_sec) as date)
cast(from_unixtime(unixtime_milli) as date)
date_format(from_unixtime(timestamp_sec), '%Y-%m-%dT%h:%i:%s')
date_format(from_unixtime(timestamp_milli), '%Y-%m-%dT%h:%i:%s')
from_iso8601_timestamp(datetime)
What are the offset and time expression parameters that you are using?
Since delta windows are effectively filters inserted into your SQL, you can troubleshoot them by manually inserting the filter expression into your data set's query.
Namely, applying a delta window filter with -3 minute (negative) offset and 'from_unixtime(my_timestamp)' time expression to a 'SELECT my_field FROM my_datastore' query translates to an equivalent query:
SELECT my_field FROM
(SELECT * FROM "my_datastore" WHERE
(__dt between date_trunc('day', iota_latest_succeeded_schedule_time() - interval '1' day)
and date_trunc('day', iota_current_schedule_time() + interval '1' day)) AND
iota_latest_succeeded_schedule_time() - interval '3' minute < from_unixtime(my_timestamp) AND
from_unixtime(my_timestamp) <= iota_current_schedule_time() - interval '3' minute)
Try using a similar query (with no delta time filter) with correct values for offset and time expression and see what you get, The (_dt between ...) is just an optimization for limiting the scanned partitions. You can remove it for the purposes of troubleshooting.
Please try the following:
Set query to SELECT * FROM dev_iot_analytics_datastore
Data selection filter:
Data selection window: Delta time
Offset: -1 Hours
Timestamp expression: from_unixtime(timestamp_sec)
Wait for dataset content to run for a bit, say 15 minutes or more.
Check contents
After several weeks of testing and trying all the suggestions in this post along with many more it appears that the extremely technical answer was to 'switch off and back on'. I deleted the whole analytics stack and rebuild everything with different names and it now seems to now be working!
Its important that even though I have flagged this as the correct answer due to the actual resolution. Both the answers provided by #Populus and #Roger are correct had my deployment being functioning as expected.
I found by chance that changing SELECT * FROM datastore to SELECT id1, id2, ... FROM datastore solved the problem.