BSP vs Non-migrated

Active customers that are not migrated to treble BSP

Impact

This notebook illustrates the use of Seasonal-Trend decomposition for a monthly time series built from total deployments of non-migrated customers (mostly Sanuker) vs Cloud customers. We are mostly interested in observing the trend of both time series, which represents the long term change, and detect specific behaviors based on historical data.

# !pip install statsmodels import numpy as np import matplotlib.pyplot as plt from scipy.optimize import curve_fit from statsmodels.tsa.seasonal import STL, seasonal_decompose

-- Non-migrated customers monthly deployments SELECT COUNT(*) as total_deployments, DATE_TRUNC('MONTH', survey_user.created_at) as month from main_public.survey_user join main_public.channel on survey_user.channel_id = channel.id where survey_user.created_at BETWEEN '2021-06-01' AND '2023-06-01' and is_valid = 't' and parent_survey_user is null and channel_provider != 'CLOUD' AND channel.hidden = 'f' group by month ORDER BY month ASC;

-- Cloud Customers monthly deployments SELECT COUNT(*) as total_deployments, DATE_TRUNC('MONTH', survey_user.created_at) as month from main_public.survey_user join main_public.channel on survey_user.channel_id = channel.id where survey_user.created_at BETWEEN '2021-06-01' AND '2023-06-01' and is_valid = 't' and parent_survey_user is null and channel_provider = 'CLOUD' AND channel.hidden = 'f' group by month ORDER BY month ASC;

-- Cloud Customers sans Rappi SELECT COUNT(*) as total_deployments, DATE_TRUNC('MONTH', survey_user.created_at) as month from main_public.survey_user join main_public.channel on survey_user.channel_id = channel.id where survey_user.created_at BETWEEN '2021-06-01' AND '2023-06-01' and is_valid = 't' and parent_survey_user is null and channel_provider = 'CLOUD' AND channel.hidden = 'f' and channel.company_id NOT IN ( SELECT id FROM main_public.company WHERE big_company = 't' and name LIKE '%Rappi%' ) group by month ORDER BY month ASC;

# Plot data in semilog scale conv_from_non_migrated_channels.set_index('month', inplace=True, drop=True) conv_from_cloud_channels.set_index('month', inplace=True, drop=True) conv_from_cloud_non_rappi.set_index('month', inplace=True, drop=True) fig, ax = plt.subplots() ax.grid(True) plt.plot(conv_from_non_migrated_channels, label='Non-migrated') plt.plot(conv_from_cloud_channels, label='Cloud') plt.plot(conv_from_cloud_non_rappi, label='Non-rappi') plt.yscale("log") plt.xticks(rotation = 45) plt.legend() plt.show()

From the plot above, we can observe that traffic from non-migrated customers is almost two orders of magnitud smaller than Cloud customers. When removing from Cloud customers those company ids matching Rappi accounts, traffic seems to be almost one order of magnitude larger than non-migrated customers. It is known that Sanuker and platforms other than Cloud do not perform well, and they could potentially from migrating to Cloud. However, it does not seem the number of deployments generated by those customers is just limited by the platform, but also but their size or needs. Let's see what the trend analysis shows.

# Timeseries decomposition for non-migrated customers # Compute over period of 6 months, smooth over period of 7 months nonmigrated_stl = STL(conv_from_non_migrated_channels.total_deployments, period=6, trend=7) nonmigrated_tsa = nonmigrated_stl.fit() fig, ax = plt.subplots() ax.grid(True) plt.xticks(rotation = 45) plt.plot(conv_from_non_migrated_channels.index, conv_from_non_migrated_channels.total_deployments, c='blue', label='Non-migrated') plt.plot(nonmigrated_tsa.trend.index, nonmigrated_tsa.trend, c='red', label='Trend') plt.legend() plt.show()

We can clearly observe a pronounced negative trend in the plot above, which is a sign of a steady decrease in the use of deployments. Now let's take a look to the Cloud customers time series.

# Timeseries decomposition for cloud customers cloud_stl = STL(conv_from_cloud_channels.total_deployments, period=6, trend=7) cloud_tsa = cloud_stl.fit() fig, ax = plt.subplots() ax.grid(True) plt.xticks(rotation = 45) plt.plot(conv_from_cloud_channels.index, conv_from_cloud_channels.total_deployments, c='blue', label='Cloud') plt.plot(cloud_tsa.trend.index, cloud_tsa.trend, c='red', label='Trend') plt.legend() plt.show()

# Timeseries decomposition for non-rappi cloud customers nonrappi_stl = STL(conv_from_cloud_non_rappi.total_deployments, period=6, trend=7) nonrappi_tsa = nonrappi_stl.fit() fig, ax = plt.subplots() ax.grid(True) plt.xticks(rotation=45) plt.plot(conv_from_cloud_non_rappi.index, conv_from_cloud_non_rappi.total_deployments, c='blue', label='Cloud non-rappi') plt.plot(nonrappi_tsa.trend.index, nonrappi_tsa.trend, c='red', label='Trend') plt.legend() plt.show()

Note: We can observe negative values for trend at the beginning of the period, due to the way statsmodels fit the curve.

As opposed as the non-migrated trend plot, Cloud and Cloud non-rappi plots show a positive trend, which is explained by various factors: 1) All new customers are provisioned in Cloud, 2) Cloud API allows for more deployments, and 3) Cloud API is generally more available (less deployments are lost).

dydx = nonrappi_tsa.trend / nonmigrated_tsa.trend fig, ax = plt.subplots() ax.grid(True) plt.xticks(rotation=45) plt.plot(conv_from_cloud_non_rappi.index, dydx, c='blue', label='dy/dx') plt.legend() plt.show()

Think of dy/dx as the rate of change of BSP non-rappi customers trend with respect to the non-migrated customers trend. Since trend is already a smoothed first-order difference, think of dy/dx as the acceleration of deployments in time. The plot above shows how BSP deployments are currently accelerating at a rate of 30x with respect to non-migrated deployments. However, the time series is relatively short (only 2 years) and this statement might not be conclusive.

In conclusion, non-migrated customers are generating a relatively marginal traffic with respect to Cloud customers, even when excluding Rappi, plus their deployments are steadily decreasing. If the effort of migrating those customers is too high (because involves a lot of manual work), it is recommended to communicate when those platforms are being sunset, and migrated them on-demand, since they don't contribute much to the total traffic (and will contribute even less in the near term).