convex-best-practices

Convex best practices

The patterns that keep a Convex app fast and correct in production. The rule that matters most: read as little as possible before you write, and read through an index.

Rules that matter most

Validators on every function. args and returns, with returns: v.null() when nothing comes back.
Indexes, not filters. Every table read goes through withIndex against an index in convex/schema.ts.
Idempotent mutations. Return early when the document is already in the target state so retries are safe.
Patch without reading first. ctx.db.patch(id, fields) throws if the doc is missing; you rarely need the old value.
Promise.all for independent writes. Do not await them one at a time.
Schedule internal.* only. Crons and ctx.scheduler run without a client, so public targets skip auth.
Thin wrappers. Auth and business logic live in plain helpers that take ctx.
ConvexError for anything a client should read. Plain Error messages are redacted in production.

// convex/tasks.ts
import { query, mutation } from "./_generated/server";
import { v, ConvexError } from "convex/values";

const taskValidator = v.object({
  _id: v.id("tasks"),
  _creationTime: v.number(),
  userId: v.id("users"),
  title: v.string(),
  status: v.union(v.literal("open"), v.literal("done")),
});

export const listOpen = query({
  args: { userId: v.id("users") },
  returns: v.array(taskValidator),
  handler: async (ctx, args) => {
    return await ctx.db
      .query("tasks")
      .withIndex("by_user_and_status", (q) =>
        q.eq("userId", args.userId).eq("status", "open"),
      )
      .order("desc")
      .take(100);
  },
});

export const rename = mutation({
  args: { taskId: v.id("tasks"), title: v.string() },
  returns: v.null(),
  handler: async (ctx, args) => {
    if (args.title.trim().length === 0) {
      throw new ConvexError("Title cannot be empty");
    }
    await ctx.db.patch(args.taskId, { title: args.title.trim() });
    return null;
  },
});

The schema behind that index:

tasks: defineTable({
  userId: v.id("users"),
  title: v.string(),
  status: v.union(v.literal("open"), v.literal("done")),
})
  .index("by_user", ["userId"])
  .index("by_user_and_status", ["userId", "status"]),

Name indexes after their fields in order, and query fields in that same order.

OCC and write conflicts

Convex runs mutations under optimistic concurrency control. A mutation records what it read. If another mutation commits a change to any of that data first, Convex retries it. After enough retries it fails and the client sees a write conflict error.

Conflicts come from three places:

Two mutations writing the same document at once: counters, "last seen" fields, a shared settings doc
A mutation that reads a wide range, such as .collect() on a whole table, so any change in that range conflicts with it
A client calling the same mutation faster than it can commit: typing, dragging, polling

Idempotent and patch first

export const complete = mutation({
  args: { taskId: v.id("tasks") },
  returns: v.null(),
  handler: async (ctx, args) => {
    const task = await ctx.db.get(args.taskId);
    if (!task || task.status === "done") {
      return null;
    }
    await ctx.db.patch(args.taskId, { status: "done" });
    return null;
  },
});

export const reorder = mutation({
  args: { itemIds: v.array(v.id("items")) },
  returns: v.null(),
  handler: async (ctx, args) => {
    await Promise.all(
      args.itemIds.map((id, index) => ctx.db.patch(id, { order: index })),
    );
    return null;
  },
});

The read in complete is fine: one document, early exit. reorder never reads at all.

Event records instead of counters

A counter field on one document is the most common conflict source. Insert one row per event and count in a query.

export const trackView = mutation({
  args: { pageId: v.id("pages") },
  returns: v.null(),
  handler: async (ctx, args) => {
    await ctx.db.insert("pageViews", { pageId: args.pageId });
    return null;
  },
});

export const viewCount = query({
  args: { pageId: v.id("pages") },
  returns: v.number(),
  handler: async (ctx, args) => {
    const views = await ctx.db
      .query("pageViews")
      .withIndex("by_page", (q) => q.eq("pageId", args.pageId))
      .collect();
    return views.length;
  },
});

Once the event table gets large, move the count to the @convex-dev/sharded-counter or @convex-dev/aggregate component instead of collecting rows.

Dedup windows

For heartbeats and presence, skip the write when the last one was recent. Pair it with a client side debounce: 300 to 500 ms for typing, a few seconds for heartbeats.

const DEDUP_MS = 10_000;

export const heartbeat = mutation({
  args: { sessionId: v.string(), path: v.string() },
  returns: v.null(),
  handler: async (ctx, args) => {
    const now = Date.now();
    const existing = await ctx.db
      .query("sessions")
      .withIndex("by_session", (q) => q.eq("sessionId", args.sessionId))
      .unique();
    if (!existing) {
      await ctx.db.insert("sessions", { ...args, lastSeen: now });
      return null;
    }
    if (existing.path === args.path && now - existing.lastSeen < DEDUP_MS) {
      return null;
    }
    await ctx.db.patch(existing._id, { path: args.path, lastSeen: now });
    return null;
  },
});

Put hot fields such as lastSeen in their own table so those writes do not conflict with reads of the stable document.

Pagination over collect

.collect() on an unbounded table gets slower every day and eventually hits read limits. Paginate anything a user can grow. postValidator below is a hoisted document validator like taskValidator above.

import { paginationOptsValidator } from "convex/server";

export const feed = query({
  args: { userId: v.id("users"), paginationOpts: paginationOptsValidator },
  returns: v.object({
    page: v.array(postValidator),
    isDone: v.boolean(),
    continueCursor: v.string(),
    splitCursor: v.optional(v.union(v.string(), v.null())),
    pageStatus: v.optional(
      v.union(v.literal("SplitRecommended"), v.literal("SplitRequired"), v.null()),
    ),
  }),
  handler: async (ctx, args) => {
    return await ctx.db
      .query("posts")
      .withIndex("by_user", (q) => q.eq("userId", args.userId))
      .order("desc")
      .paginate(args.paginationOpts);
  },
});

On the client, usePaginatedQuery from convex/react drives loadMore.

No Date.now() in queries

Queries must be deterministic so Convex can cache them and rerun them when data changes. Pass time from the client, or store a status field that a scheduled mutation updates.

export const dueBefore = query({
  args: { userId: v.id("users"), now: v.number() },
  returns: v.array(taskValidator),
  handler: async (ctx, args) => {
    return await ctx.db
      .query("tasks")
      .withIndex("by_user_and_due", (q) =>
        q.eq("userId", args.userId).lt("dueAt", args.now),
      )
      .take(100);
  },
});

Mutations and actions may call Date.now().

ESLint plugin

@convex-dev/eslint-plugin catches the old function syntax, missing arg validators, .filter() in queries, and top of the hour crons at lint time. Install it in every Convex project.

npm i --save-dev @convex-dev/eslint-plugin

// eslint.config.js
import { defineConfig } from "eslint/config";
import convexPlugin from "@convex-dev/eslint-plugin";

export default defineConfig([...convexPlugin.configs.recommended]);

Open references/eslint-setup.md when you need the full rule list, the TypeScript aware config, package scripts, or a custom convex/ directory.

Common mistakes

Mistake	Why it breaks	Do instead
`.filter()` on a table query	Reads every row, then drops most	Add an index, use `withIndex`
`.collect()` on an unbounded table	Slower every day, hits read limits, wide OCC footprint	`.paginate()` or `.take(n)`
Read, compute, then patch a shared doc	Two clients read the same version and both write	Patch directly, or split into event rows
Counter field incremented per event	Every increment conflicts with every other	Event records or a sharded counter component
Mutation without an early return	Retries and double clicks apply the change twice	Check state, return `null` if already done
Sequential `await` on independent writes	Slow, and each read widens the conflict window	`Promise.all`
`Date.now()` in a query	Result changes every ms, cache and subscriptions break	Pass `now` as an arg
Scheduling `api.*`	Public function runs with no client auth	Schedule `internal.*`
Plain `Error` for user messages	Redacted to "Server Error" in production	`ConvexError`
Business logic inside the handler	Untestable, duplicated across functions	Plain helper that takes `ctx`