Data Visualization

The Object-Oriented API

Every plot belongs to a Figure containing Axes — grab them explicitly:

import matplotlib.pyplot as plt

fig, ax = plt.subplots(figsize=(8, 4))
ax.plot(x, y, label="throughput", color="#1f77b4", linestyle="--")
ax.scatter(x2, y2, marker="o", s=40)
ax.set_xlabel("requests/s"); ax.set_ylabel("p99 ms")
ax.set_title("Latency under load")
ax.legend(); ax.grid(True, alpha=0.3)
fig.tight_layout()

The stateful shorthand (plt.plot) targets the "current" axes — convenient in a REPL, error-prone when building multiple panels because a stray call can draw into the wrong subplot. The OO form scales from one line to dashboards: ax is a handle you can pass into helper functions that add layers without global state.

Panels, Ticks, and Saving

fig, axes = plt.subplots(2, 2, figsize=(10, 7), sharex=True)
for a, (title, frame) in zip(axes.flat, frames):
    a.hist(frame["value"], bins=30)
    a.set_title(title)

ax.set_xticks([0, 5, 10])
ax.set_xticklabels(["low", "mid", "high"], rotation=30)
fig.savefig("report.png", dpi=150, bbox_inches="tight")

bbox_inches='tight' crops the margin whitespace that otherwise floats labels off-canvas; dpi controls print resolution. For dates on the x-axis, pass datetime objects directly and call fig.autofmt_xdate(). Headless servers should select a non-GUI backend before importing pyplot: matplotlib.use("Agg") or MPLBACKEND=Agg — plt.show() then becomes a no-op and savefig does the work.

Seaborn Layering and Pitfalls

Seaborn is a layer on matplotlib, not a replacement — it returns the same Figure/Axes objects:

import seaborn as sns
sns.set_theme(style="darkgrid", context="talk")   # styling + scaled fonts

g = sns.relplot(data=df, x="hour", y="orders",
                hue="region", kind="line", col="year")
sns.regplot(data=df, x="temp", y="sales", ax=ax)   # drop into an existing Axes

Common problems:

• savefig after show may save a cleared figure — save first
• Open figures leak memory in loops: call plt.close(fig)
• Categorical axis order follows first appearance — set hue_order/order to control it
• Global style: call sns.set_theme() once at entry, not per plot

Interview angle: describe a layered workflow — data → seaborn for statistical geometry → matplotlib ax for annotation and tweaks — and the CI backend choice.