WITH daily_users AS (
    SELECT 
        event_date,
        user_id
    FROM user_events
    WHERE event_date BETWEEN '2024-01-01' AND '2024-01-30'
),
date_series AS (
    SELECT DISTINCT event_date 
    FROM daily_users
),
cumulative_users AS (
    SELECT 
        d.event_date,
        COUNT(DISTINCT u.user_id) AS cumulative_unique_users
    FROM date_series d
    LEFT JOIN daily_users u ON u.event_date <= d.event_date
    GROUP BY d.event_date
    ORDER BY d.event_date
)
SELECT * FROM cumulative_users;

WITH daily_feature_users AS (
    SELECT 
        event_date,
        user_id
    FROM user_events
    WHERE event_name = 'feature_x' 
      AND event_date BETWEEN '2024-01-01' AND '2024-01-30'
),
daily_total_users AS (
    SELECT 
        event_date,
        user_id
    FROM user_events
    WHERE event_date BETWEEN '2024-01-01' AND '2024-01-30'
),
date_series AS (
    SELECT DISTINCT event_date 
    FROM daily_total_users
),
cumulative_feature_users AS (
    SELECT 
        d.event_date,
        COUNT(DISTINCT u.user_id) AS cumulative_feature_users
    FROM date_series d
    LEFT JOIN daily_feature_users u ON u.event_date <= d.event_date
    GROUP BY d.event_date
    ORDER BY d.event_date
),
cumulative_total_users AS (
    SELECT 
        d.event_date,
        COUNT(DISTINCT u.user_id) AS cumulative_total_users
    FROM date_series d
    LEFT JOIN daily_total_users u ON u.event_date <= d.event_date
    GROUP BY d.event_date
    ORDER BY d.event_date
)
SELECT 
    cfu.event_date,
    cfu.cumulative_feature_users,
    ctu.cumulative_total_users,
    ROUND(100.0 * cfu.cumulative_feature_users / (ctu.cumulative_total_users, 0), 2) AS penetration_rate
FROM cumulative_feature_users cfu
JOIN cumulative_total_users ctu ON cfu.event_date = ctu.event_date
ORDER BY cfu.event_date;

“А это для выгрузок или для ежедневного контроля?”

“Нам вообще-то важен срез по регионам, почему его нет? А по приложениям? Точкам входа в продукт?”

“А где сравнение с прошлым месяцем?”

"А нам еще важно следить за этим, сделаешь парочку графиков?

"Смежникам важно смотреть на это, а давай еще сделаем так: на одной вкладке будет общая информация, а на второй возможность для финансов грепнуть

Метрики перформанса могут влиять и на продуктовые показатели. К такому выводу мы пришли после проведения ухудшающих тестов.

If it helps in the short-term, but hurts Duolingo in the long-term, it’s not right

When we disagree, we test ideas and let the metrics decide.

Learners need to get utility out of whatever we’ve built. Otherwise, we’ve made something that adds more complexity to the app and distracts learners from what they’re here to do.

Every feature we ship must be intuitive, delightful, useful, and polished.

If a feature or screen requires explanation or additional context, it’s not right.

Some of our best features and campaigns have come from asking ridiculous, unlikely questions

Talking about ideas is rarely as effective as building them

We don’t chase blame. We dig in and figure out what happened. This is how we continuously raise the bar—not by being perfect, but by being a little bit better each day.

from sklearn.linear_model import LinearRegression
from sklearn.datasets import make_regression
import pandas as pd

X, y = make_regression(n_samples=1000, n_features=10, noise=10, random_state=42)
feature_names = [f"feature_{i}" for i in range(X.shape[1])]

model = LinearRegression()
model.fit(X, y)

importance = pd.Series(model.coef_, index=feature_names).sort_values(ascending=False)

from catboost import CatBoostRegressor

model = CatBoostRegressor(iterations=200, verbose=False)
model.fit(X, y)

feature_importance = pd.Series(model.get_feature_importance(), index=feature_names)

from sklearn.ensemble import RandomForestRegressor
from sklearn.inspection import permutation_importance
from sklearn.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

model = RandomForestRegressor(n_estimators=100)
model.fit(X_train, y_train)

perm_importance = permutation_importance(model, X_test, y_test, scoring="r2")
importance = pd.Series(perm_importance.importances_mean, index=feature_names)

import shap

explainer = shap.TreeExplainer(model)
shap_values = explainer.shap_values(X_test)

shap.summary_plot(shap_values, X_test, feature_names=feature_names)

Линейная регрессия подходит для простых зависимостей, деревья и бустинг – для сложных, Permutation Importance – когда нужна честная оценка вклада признаков, а SHAP – для объяснения модели

WITH old_product AS (
    SELECT 
        user_id, 
        COUNT(order_id) AS old_orders, 
        SUM(revenue) AS old_revenue
    FROM orders
    WHERE product_id = 'old_product'
      AND order_date < '2024-01-01'
    GROUP BY user_id
),

new_product AS (
    SELECT 
        user_id, 
        COUNT(order_id) AS new_orders, 
        SUM(revenue) AS new_revenue
    FROM orders
    WHERE product_id = 'new_product'
      AND order_date >= '2024-01-01'
    GROUP BY user_id
)

SELECT 
    o.user_id,
    o.old_orders,
    n.new_orders,
    o.old_revenue,
    n.new_revenue,
    CASE 
        WHEN o.old_orders > 0 AND n.new_orders > 0 AND n.new_revenue < o.old_revenue 
            THEN 'Перешли, но платят меньше'
        WHEN o.old_orders > 0 AND n.new_orders > 0 AND n.new_revenue >= o.old_revenue 
            THEN 'Перешли, но выручка сохранилась'
        WHEN o.old_orders > 0 AND n.new_orders = 0 
            THEN 'Остались на старом'
        WHEN o.old_orders = 0 AND n.new_orders > 0 
            THEN 'Новый клиент'
        ELSE 'Нет активности'
    END AS user_behavior
FROM old_product o
LEFT JOIN new_product n ON o.user_id = n.user_id;